FINANCIAL RISK MANAGEMENT 2

Ta Quoc Bao

Department of Mathematics,
International University-VNUHCM

Ta Quoc Bao (IU) FRM2 1 / 56

Chapter 2. Univariate Volatility Modeling
• The volatility plays a crucial role in financial risk management, it is
the main measure of risk. On the other hands, the volatility is the
key factor in, e.g., Investment decisions, Portfolio construction
(Markowitz model) and Derivative pricing (Black-Scholes model).
• In this Chapter we focus on the estimation and forecasting of
volatility for a single asset (univariate)

Ta Quoc Bao (IU) FRM2 2 / 56

Chapter 2. Univariate Volatility Modeling
• The volatility plays a crucial role in financial risk management, it is
the main measure of risk. On the other hands, the volatility is the
key factor in, e.g., Investment decisions, Portfolio construction
(Markowitz model) and Derivative pricing (Black-Scholes model).
• In this Chapter we focus on the estimation and forecasting of
volatility for a single asset (univariate)

2.1. Stationary processes

• A time series is a sequence of observations in chronological order.
For example: daily log returns on a stock or monthly values of the
Consumer Price Index (CPI)
• A stochastic process is a sequence of random variables and can
be viewed as the “theoretical” or “population” analog of a time
series, conversely, a time series can be considered a sample from
a stochastic process.
Ta Quoc Bao (IU) FRM2 2 / 56
Denote {Xt , t ∈ I} the time series, where I is a time index. For example:
I = {1, 2, 3, ...}; I = {2000, 2001, 2002...2021}. Equally spaced time
series are the most common in practice. This is the case of
I = {t1 , t2 , ..., tn }, where
∆ = ti+1 − ti
with ∆ is a constant.
Difference from traditional Statistical Inference
• In traditional statistic inference, the data is assumed to be an i.i.d
process (random sample).
• In time series, we do not need this assumption and wish to model
the dependency among observations which leads to the concept
of autocorrelation.

Ta Quoc Bao (IU) FRM2 3 / 56

Some main problems in time series
• Formulate and estimate a parametric model for Xt (need to
propose methods of estimation and model diagnostics).
• This point is related to the estimation of autoregressive (AR) or
ARMA models.
• Estimation of Missing values (fill“gaps”).
• Prediction or Forecasting (“would like to know what a future value
is”). For example our data is x1 , x2 , ..., x100 , we wish to forecast the
next 10 values, x101 , ..., x110 . In this case, our forecasting horizon is
10.
• Plotting time series to observe fluctuations of time series, e.g., to
find stationarity or non-stationarity, cycles, trends, outliers or
interventions. Assisting in the formulation of a parametric model.

Ta Quoc Bao (IU) FRM2 4 / 56

Example
Consider Financial Index SP500. The data consists of excess returns
Xt = log(St ) − log(St−1 ). From the plot we see the following properties
of Xt
• The mean level of the process seems constant
• There are sections of the data with explosive behavior(high
volatility).
• The data corresponds to a non-stationary process. (will define
more detailed)
• The variance (or volatility) is not constant in time.
• No linear time series model will be available for this data.

Ta Quoc Bao (IU) FRM2 5 / 56

 0.10
0.05
0.00
−0.05
returns

−0.10
−0.15
−0.20

2000 4000 6000 8000

time

Ta Quoc Bao (IU) FRM2 6 / 56

Definition 1 (Autocovariance)
The autocovariance function a stochastic process X is defined as

γ(t, τ ) = E(Xt − µt )(Xt−τ − µt−τ )

for τ ∈ Z, where µt = E(Xt ).

• The autocovariance function is symmetric, i.e.,

γ(t, τ ) = γ(t − τ, −τ )

For special case τ = 0 then γ(t, 0) = Var(Xt )

• In general γ(t, τ ) is depend on t as well as τ

Example
Find autocovariance function of Brownian motion?

Ta Quoc Bao (IU) FRM2 7 / 56

Definition 2
A process is said to be strictly stationary if all aspects of its behavior
are unchanged by shifts in time. Mathematically, stationarity is defined
as the requirement that for every m and n the distribution of
(X1 , X2 , ..., Xn ) and (X1+m , X2+m , ..., Xn+m ) are the same.

Definition 3
A process is weakly stationary if its mean, variance, and covariance
are unchanged by time shifts. More precisely, X1 , X2 , ..., is a weakly
stationary process if
(i) E(Xt ) = µ for all t
(ii) Var(Xt ) = σ 2 (a positive finite constant) for all t
(iii) Cov(Xt , Xs ) = γ(| t − s |) for all t, s and some function γ
We see that, the mean and variance do not change with time and the
covariance between two observations depends only on the lag, the
time distance | t − s |.
Ta Quoc Bao (IU) FRM2 8 / 56
 • The function γ is the autocovariance function of the process and
has symmetric property

γ(h) = γ(−h)

• The correlation between Xt and Xt+h is denoted by ρ(h), function ρ

is called autocorrelation function (ACF). We have γ(0) = σ 2 and,
hence
γ(h) = σ 2 ρ(h) hence ρ(h) = γ(h)/γ(0).

The ACF is normalized on [−1, 1]. Since the process is required to be

covariance stationary, the ACF depends only on one parameter, lag h.

Ta Quoc Bao (IU) FRM2 9 / 56

Example
Consider the random walk X: Xt = c + Xt−1 + t , with c is constant and
white noise t . We see that if c 6= 0, then Zt := Xt − Xt−1 = c + t have a
non-zero mean. We call it a random walk with drift. Note that since t is
independent then we call Xt a random walk with independent
increments.
For more convenience, assume that c and X0 are set to zero. We have

Xt = t + t−1 + ... + 1

Hence
µt = E(Xt ) = 0
and
Var(Xt ) = tσ

Ta Quoc Bao (IU) FRM2 10 / 56

 γ(t, s) = Cov(Xt , Xs ) = (t − s)σ 2
If s < t then r
s
ρ(t, s) = 1 −
t
which against ρ depending on t as well as on s, thus the random walk
is not covariance stationary. The following figure shows the
relationship among different processes: Stationary processes are the
largest set, followed by white noise, martingale difference (MD), and
i.i.d. processes.

Ta Quoc Bao (IU) FRM2 11 / 56

Estimating Parameters of a Stationary Process
Let X1 , X2 , ..., Xn be observations from weakly stationary process. To
estimate the autocovariance function, we use the sample
autocovariance function defined by
n−h
1X
γ̂(h) = (Xt+h − X̄)(Xt − X̄)
n
t=1

To estimate function ρ, we use the sample autocorrelation function

(sample ACF) defined as
γ̂(h)
ρ̂(h) =
γ̂(0)

Ta Quoc Bao (IU) FRM2 12 / 56

• To visualize the dependencies of xt for different lags h, we use the
Correlogram.
• A correlogram is a plot of h (x-axis) versus its corresponding value
of ρ̂(h) (y-axis).
• The correlogram may exhibit patterns and different degrees of
dependency in a time series.
√
• A “band”of size 2/ n is added to the correlogram because
asymptotically ρ̂(h) ∼ N(0, 1/n) if the data is close to a white noise
process.
• This band is used to detect significant autocorrelations, i.e.
autocorelations that are different from zero.

Ta Quoc Bao (IU) FRM2 13 / 56

 Microsoft's return ACF

1.0
0.8
0.6
ACF

0.4
0.2
0.0

0 5 10 15 20 25 30 35

Lag

Ta Quoc Bao (IU) FRM2 14 / 56

Testing stationarity
Augmented Dickey-Fuller Test (ADF) (also called Unit Root Test)
The test uses the following null and alternative hypotheses:
• H0 : The time series is non-stationary, i.e., it has some
time-dependent structure and does not have constant variance
over time.
• H1 : The time series is stationary.
In R, use: adf.test
KPSS test
The ideas of KPSS test comes from the regression model with time
trend
t
X
Xt = c + µt + k ξi + η t
i=1

Ta Quoc Bao (IU) FRM2 15 / 56

with stationary ηt and i.i.d ξ with mean 0 and variance 1. Note that the
third term is a random walk. So we set the hypothesis and alternative
as follows
• H0 : k = 0, i.e., the test is that the data is stationary
• H1 : k 6= 0
in R, use kpss.test. Test results for Microsoft data

Ta Quoc Bao (IU) FRM2 16 / 56

Ljung–Box Test for autocorrelations
Sample ACF with test bounds.
• These bounds are used to test the null hypothesis that an
autocorrelation coefficient is 0. The null hypothesis is rejected if
the sample autocorrelation is outside the bounds.
• The usual level of the test is 0.05
Example (The First-order Autoregression Model (AR(1))) The time
series X = (Xt ) is called AR(1) if the value of X at time t is a linear
function of the value of X at time t − 1 as follows
∞
X
Xt = δ + φ1 Xt−1 + wt = .. = δ + φh1 wt−h
h=0

Ta Quoc Bao (IU) FRM2 17 / 56

where
(i) the error wt ∼ N(0, σw2 ) and are i.i.d
(ii) wt is independent of Xt
(iii) | φ1 |< 1 this condition grantees that Xt is weakly stationary
We have
• The (theoretical) mean of Xt
δ
E(Xt ) = µ =
1 − φ1
• The variance of Xt is
σw2
Var(Xt ) =
1 − φ21
σw 2
• The covariance: Cov(Xt , Xt+h ) = γ(h) = φh1 1−φ 2
1
• The correlation between observations h time periods apart is
ρ(h) = φh1

Ta Quoc Bao (IU) FRM2 18 / 56

Note that the magnitude of its ACF decays geometrically to zero, either
slowly as when φ1 = 0.95, moderately slowly as when φ1 = 0.75, or
rapidly as when φ1 = 0.25. We now simulate AR(1) and plot the ACF
with φ1 = 0.64 and σw2 = 1

Series x_t

1.0
3

0.8
2

0.6
1

ACF

0.4
xt

0.2
−1

0.0
−2

−0.2

0 20 40 60 80 100 1 2 3 4 5

Time Lag

Ta Quoc Bao (IU) FRM2 19 / 56

Ljung–Box Test. The null hypothesis of the Ljung–Box test is

H0 : ρ(1) = ρ(2) = ...ρ(m) = 0

for some m. If the Ljung–Box test rejects, then we conclude that one or
more of ρ(1), ρ(2), ..., ρ(m) is nonzero. The Ljung–Box test is
sometimes called simply the Box test.
m
X ρ̂2 (j)
Q(m) = n(n + 2) ∼ χ2 (m)
n−j
i=j

Example: Consider AR(1) with φ1 = 0.64 and σw2 = 1, we have the

results of Box test in R
Box.test(xt , lag = 10, type = "Ljung-Box")
X-squared = 50.935, df = 10, p-value = 1.796e-07

Ta Quoc Bao (IU) FRM2 20 / 56

Example (Nonstationary AR(1) Processes)
If | φ1 |>≥ 1 then AR(1) process is nonstationary, and the mean,
variance, covariances and and correlations are not constant.
• Random Walk (φ = 1)

Xt = Xt−1 + wt = X0 + w1 + w2 + ... + wt

and the process is not stationary. Hence, for all t

E(Xt | X0 ) = X0

which is constant but depends entirely on the arbitrary starting

point. Moreover,
Var(Xt ) = tσw2
which is not stationary but rather increases linearly with time and
makes the random walk “wander”, i.e.,Xt takes increasingly longer
excursions away from its conditional mean of X0 , and therefore is
not mean-reverting.
Ta Quoc Bao (IU) FRM2 21 / 56
AR, MA and ARMA models
• The autoregressive process of order p or AR(p) is defined by the
equation
Xp
Xt = φj Xt−j + ωt
j=1

where ωt ∼ N(0, σ 2 )
• The AR model establishes that a realization at time t is a linear
combination of the p previous realization plus some noise term. If
p = 0 then Xt = ωt there is no autoregression term.
• The lag operator (Back-shift operator) is denoted by B and used to
express lagged values of the process so, Bj Xt = Xt−j
• Define
p
X
Φ(B) = 1 − φj Bj
j=1

Ta Quoc Bao (IU) FRM2 22 / 56

Then we have AR(p) can be written by
Φ(B)Xt = ωt , t = 1, 2, ..., n

• Φ(B) is known as the characteristic polynomial of the process and

its roots determine when the process is stationary or not.
The moving average process of order q, denoted MA(q), defined as
q
X
Xt = ωt + θj ωt−j
j=1

• Under this model, the observed process depends on previous ωt

• MA(q) can define correlated noise structure in our data and goes
beyond the traditional assumption where errors are iid.
• We have Xt = Θ(B)(ωt ), where Θ(B) = 1 + qj=1 θj Bj
P

The general autoregressive moving average process of orders p and q

or ARMA(p, q) combines both AR(p) and MA(q) models into a unique
representation.
Ta Quoc Bao (IU) FRM2 23 / 56
Consider the AR(p). The process Xt is stationary if and only if all root of
the characteristic equation
p
X
Φ(α) = 1 − φj α j = 0
j=1

are greater than one, i.e., | αj |> 1 for all j = 1, 2, .., p

Ta Quoc Bao (IU) FRM2 24 / 56

Partial Autocorrelation Function (PACF)
A partial correlation is a conditional correlation. It is the correlation
between two variables under the assumption that we know and take
into account the values of some other set of variables.
• The partial autocorrelation function (PACF) of a process Xt is
defined as

pk = Corr(Xt , Xt+k | Xt+1 , Xt+2 ..., Xt+k−1 ), k = 0, 1, 2, ..

• The PACF can also be derived through an autoregressive model

of order k

Xt+k = φk1 Xt+k−1 + φk2 Xt+k−2 + ... + φkk Xt + ωt+k

ωt+k is normal error term uncorrelated with Xt+k−j , j ≥ 1

• The coefficients φk1 , φk2 , ..., φkk define the PACF
• In R we use: acf(data, type=’partial’)
Ta Quoc Bao (IU) FRM2 25 / 56
• Suppose that Xt is zero mean stationary process.
• Multiplying Xt+k−j on both sides of the above regression equation
and taking the expectation, we get

γj = φk1 γj−1 + φk2 γj−2 + ... + φkk γj−k

• If we divide by γ0 we get,

ρj = φk1 ρj−1 + φk2 ρj−2 + ... + φkk ρj−k

• Now for j = 1, 2, ..., k, we have the following system of equations

with variables:

ρ1 = φk1 ρ0 + φk2 ρ1 + ... + φkk ρk−1

ρ2 = φk1 ρ1 + φk2 ρ0 + ... + φkk ρk−2
..
.
ρk = φk1 ρk−1 + φk2 ρk−2 + ... + φkk ρ0

Ta Quoc Bao (IU) FRM2 26 / 56

• We can write the system of equations as

Pk φk = ρ

where  
1 ρ1 · · · ρk−1
 ρ1 1 · · · ρk−1 
Pk = 
 
.. .. .. .. 
 . . . . 
ρk−1 ρk−2 ··· 1
φk = (φk1 , φk2,...,φkk )T and ρ = (ρ1 , ρ2 , ..., ρk ). We get

| P∗k |
φkk =
| Pk |

where P∗k is is matrix obtained from Pk by replaced k−th with ρ.

Ta Quoc Bao (IU) FRM2 27 / 56

 • For PACF and ACF order 1, there is no difference, we have

φ11 = ρ1

• For order 2 we have

ρ2 − ρ21
φ22 =
1 − ρ21
Example: The AR(1), process Xt = αXt−1 + ωt has ACF ρτ = ατ . For
the PACF, we have φ11 = ρ1 = α and

α2 − α2
φ22 = =0
1 − α2
and φkk = 0 for all k > 1.
In general, for AR(p), we have ρkk = 0 for all k > p

Ta Quoc Bao (IU) FRM2 28 / 56

Example: Consider MA(1); process Xt = βωt−1 + ωt , with Var(ωt = σ 2 .
β
It holds that γ0 = σ 2 (1 + β 2 ), ρ1 = 1+β 2 , and ρk = 0 for all k > 1. We

have
β
φ11 = ρ1 =
1 + β2
and
ρ21
φ22 = −
1 − ρ21
For MA(1) process, it strictly holds that φ22 < 0. If we continue to
calculate with k > 2 we could determine that PACF will not reach zero.

Ta Quoc Bao (IU) FRM2 29 / 56

2.2 Volatility models
• Volatility plays important role in modelling financial system, it is
directly used to measure risk.
• Stocks, exchange rates, interest rates and other financial time
series have some stylized facts that are difference from other time
series. Stylized facts, generally speaking, are the result of many
independent empirical studies statistical properties of financial
markets that have been proven to be common across financial
markets.
• These stylized facts are arised from Effective Market Hypothesis
(EMH).
• A good candidate for modelling of financial time series should
represent the properties of stochastic process. e.g., AR or ARMA
can fulfill this task.
• Some volatility models, e.g., ARCH or GARCH can replicate these
stylized facts appropriately.
Ta Quoc Bao (IU) FRM2 30 / 56
• The efficient markets hypothesis (EMH) in finance assumes that
asset prices are fair, information is accessible for everybody and is
assimilated rapidly to adjust prices, and people (including traders)
are rational.
• The price change Pt − Pt−1 is only due to the arrival of “news”
between tand t + 1. Hence individuals have no opportunities for
making an investment with return greater than a fair payment for
undertaking riskiness of the asset. i.e., the price is right, and there
exist no arbitrage opportunities. This is called strong form the
EMH
• A semi-strong form states that security prices reflect efficiently all
public information, leaving rooms for the value of private
information. The weak form merely assumes security prices reflect
all past publicly available information.

Ta Quoc Bao (IU) FRM2 31 / 56

• Stylized fact 1. Time series of share prices Pt and other basic
financial instruments are not stationary time series and possess a
local trend at the least.
• Stylized fact 2. Returns rt have a leptokurtic distribution. The
empirically estimated kurtosis is mostly greater than 3.
• Stylized fact 3. The return process is white noise since the
sampleautocorrelation ρ̂k,n , k 6= 0 is not significantly different from
0. Furthermore the white noise is not independent since the
sample autocorrelations of squared and absolute returns are
clearly greater than 0.
• Stylized fact 4. Volatility tends to form clusters: After a large
(small) price change (positive or negative) a large (small) price
change tends to occur. This effect is called volatility clustering.

Ta Quoc Bao (IU) FRM2 32 / 56

• Under the EHM, an asset return process may be expressed as

rt = µt + t

where µt = E(rt | Ft−1 ) is the conditional mean or the rational

expectation of log-return, and t ∼ (0, σt2 ), more precisely,
Var(t | Ft−1 ) = Var(rt | Ft−1 ) = σt2 is the conditional variance.
Note that µt and σt2 are known at time t − 1.
• Note that Ft−1 = σ(r0 , r1 , .., rt−1 ). So µt usually can be followed
ARMA model. The term t is the innovation or random component
of the log-return and in practice, the conditional variance of this
innovation, σt2 , is time varying and stochastic. So EWMA,
ARCH/GARCH models may be used to model this dynamic
behavior of conditional variances.

Ta Quoc Bao (IU) FRM2 33 / 56

2.2. Exponentially weighted moving average (EWMA)
Denote yt the return of stock at time t. Then
• Volatility a weighted sum of past returns, with weights ωi , is
defined by
σ̂t2 = ω1 y2t−1 + ω2 y2t−2 + ... + ωL y2t−L
where L is the length of the estimation window, i.e., the number of
observations used in the calculation. This is called MA model
• An extension of MA model is Exponentially weighted moving
average. Let the weights be exponentially declining, and denote
them by λi
L
1−λ 1−λ X i 2
σ̂t2 = L
(λy2t−1 + λ2 y2t−2 + ... + λL y2t−L ) = λ yt−i
λ(1 − λ ) λ(1 − λL )
i=1

where 0 < λ < 1 is the decay factor. If L is large enough, the term
αn are negligible for all n > L. So we set L = ∞
Ta Quoc Bao (IU) FRM2 34 / 56
Note that the sum of weights is
∞
λ X
= λi
1−λ
i=1

So the exponentially weighted moving average is defined by

∞
1−λX 2
σ̂t2 = λi yt−i
λ
i=1

and, hence, we get the EWMA equation (why???)

σ̂t2 = λσ̂t−1
2
+ (1 − λ)y2t−1 (1)

Note that JP Morgan set for daily data with λ = 0.94

Ta Quoc Bao (IU) FRM2 35 / 56

Example
Suppose that λ = 0.9, the volatility estimated for a market variable for
day n − 1 is 1% per day, and during day n − 1 the market variable
increased by 2%. This means that σn−1 2 = 0.012 = 0.0001 and
y2n−1 = 0.022 = 0.0004. From equation (1) we get

σn2 = 0.9 × 0.0001 + 0.1 × 0.0004 = 0.00013

√
The estimate of the volatility for day n is σn = 0.00013 = 1.4% per
day. Note that the expected value of y2n−1 is σn−12 = 0.0001. Hence,
2
realized value of yn−1 = 0.0002 is greater that expected value, and as a
result our volatility estimate increase. If the realized value of y2n−1 has
been less than its expected valued, our estimate of the volatility would
have decreased.

Ta Quoc Bao (IU) FRM2 36 / 56

2.3 The ARCH and GARCH models
ARCH model.
• The ARCH model was proposed by Robert Engle in 1982 called
autoregressive conditionally heteroscadastic
• most volatility models derive from this
• Returns are assumed to have conditional distribution (here
assumed to be normal)

yt ∼ N(0, σt2 )

or we can write
yt = σt t
where t ∼ N(0, 1) is called residual. Note that in general we do
not have this assumption, i.e., Yt ∼ (0, σt2 ) and  ∼ (0, 1)

Ta Quoc Bao (IU) FRM2 37 / 56

 • ARCH(L1 ) is defined by
L1
X
Var(yt | yt−1 , yt−2 , ..., yt−L1 ) = σt2 = ω + αi y2t−i
i=1

where L1 is called the lag of the model. It is seen that in the ARCH
model, the volatility is weighted average of past returns. The most
common form is ARCH (1)

Var(yt | yt−1 ) = σt2 = ω + αy2t−1

where ω and α are parameters that can be estimated by maximum

likelihood
Note that the conditions for parameters are: ω > 0 and Li=1
P 1
αi < 1.
The second condition ensures covariance stationary.

Ta Quoc Bao (IU) FRM2 38 / 56

ARCH(1) unconditional volatility
ω
In this case we have: σ 2 = E(y2t ) = E(ω + αy2t−1 ). Hence σ 2 = 1−α

ARCH(1) with fat tail

In many cases we assume that the residuals t ∼ N(0, 1). However,
unconditional distribution of the returns may be fat. We have

E(Y 4 ) = E(Yt4 ) = E(σt4 4t ) = 3E(σt4 )

and
E(Y 4 ) = 3E((ω + αYt−1
2
)2 ) = 3ω 2 + 2αωσ + α2 E(Yt4 )
3ω(1−α)
So E(Y 4 ) = (1−α)(1−3α2 )
and the Kurtosis is:

3(1 − α2 )
Kur = > 3 ⇒ 3α2 < 1
1 − 3α2

Ta Quoc Bao (IU) FRM2 39 / 56

If we assume that the series has mean = 0 (this can always be done by
centering), then the ARCH model could be written as

yt = σt t
q
with σt = ω + αy2t−1
and t ∼ N(0, 1), i.i.d

We require that ω, α > 0 so that ω + αy2t−1 > 0 for all t. We also require
that α < 1 in order to the process to be stationary with a finite variance.
Now we have
y2t = 2t (ω + αy2t−1 )
which is similar to an AR(1) for variable y2t and with multiplicative noise
with a mean of 1 rather than additive noise with a mean of 0.

Ta Quoc Bao (IU) FRM2 40 / 56

Generalized ARCH (GARCH) model.
• It turns out that ARCH model is not a very good model and almost
nobody uses it. Because, it needs to use information from many
days before t to calculate volatility on day t. That is, it needs a lot
of lags
• The GARCH (L1 , L2 ) model is defines as
L1
X L2
X
σt2 =ω+ αi y2t−i + 2
βi σt−i
i=1 i=1

and, hence, GARCH (1,1)

σt2 = ω + αy2t−1 + βσt−1

2

• GARCH(1,1) is the most common specification

Ta Quoc Bao (IU) FRM2 41 / 56

GARCH (1,1) unconditional volatility
• The unconditional volatility (so-called the long-run variance rate) is
the unconditional expectation of volatility on given time

σ 2 = E(σt2 )

so we have

σ 2 = E(ω + αy2t−1 + βσt−1

2
) = ω + ασ 2 + βσ 2

Hence,
ω
σ2 =
1−α−β

• So to ensure positive volatility forecasts we need the condition

ω, α, β ≥ 0

Because if any parameter is negative σt+1 may be negative

Ta Quoc Bao (IU) FRM2 42 / 56
 • For stationary we need condition

α+β <1

Setting γ := 1 − α − β and V := σ 2 (called long-run variance rate). We

have
σt2 = γV + αy2t−1 + βσt−1
2

Meaning of Parameters in GARCH model

• The parameter α is news, it shows that how the volatility reacts to
new information
• The parameter β is memory, it shows that how much volatility
remembers from the past
• The sum α + β determines how quickly the predictability (memory)
of the process dies out:
• if α + β ≈ 0 predictability will die out very quickly,
• if α + β ≈ 1 predictability will die out very slowly
Ta Quoc Bao (IU) FRM2 43 / 56
Example
Suppose that a GARCH(1,1) model is estimated from daily data is

σn = 0.000002 + 0.13y2n−1 + 0.86σn−1

2

This corresponds to ω = 0.000002, α = 0.13, β = 0.86. We have

ω
σ2 = = 0.0002
1−α−β
√
or σ = 0.0002 = 0.014 = 1.4% per day.
Suppose that the estimate of the volatility on day n − 1 is 1.6% per day
2
so that σn−1 = 0.0162 = 0.000256, and on that day n − 1 the market
variable decreased by 1% so that y2n−1 = 0.012 = 0.0001. Then

σn2 = 0.000002 + 0.13 × 0.0001 + 0.86 × 0.000256 = 0.00023516

√
the new estimate of the volatility is: 0.00023516 = 0.0153 or 1.53% per
day.
Ta Quoc Bao (IU) FRM2 44 / 56
2.4. Maximum likelihood
Maximum likelihood is the most important and widespread method of
estimation. What is maximum likelihood?
• Ask the question which parameters most likely generated the data
we have
• Suppose we have a sample of

{−0.2, 3, 4, −1, 0.5}

• in the following three possibilities, which is most likely for

parameters?
case µ σ
1 1 5
2 -2 2
3 1 2

Ta Quoc Bao (IU) FRM2 45 / 56

Let Y = (y1 , y2 , ..., yn ) be a vector of data and let θ = (θ1 , θ2 , ..., θp ) be a
vector of parameters. Let f (Y | θ) be the density of Y which depends on
the parameters. The function

L(θ) := f (Y | θ)

is viewed as the function of θ with Y fixed at the observed data is called

the likelihood function.
• The maximum likelihood estimator (MLE) is the value of θ that
maximizes the likelihood function. We denote the MLE by θ̂ML .
• it is mathematically easier to maximize log L(θ), which is called the
log-likelihood. If the data are independent, then the likelihood is
the product of the marginal densities

Ta Quoc Bao (IU) FRM2 46 / 56

Application to ARCH(1)
Consider ARCH(1) model:

yt = σt t
σt2 = ω + αy2t−1
t ∼ N(0, 1)

For t = 2 we have the density??

1  1 y2 
2
f (y2 | y1 ) = q exp −
2
2π(ω + αy1 ) 2 ω + αy21

Hence, the joint density

T T
Y Y 1  1 y2t 
f (yt | yt−1 ) = exp −
2 ω + αy2t−1
q
t=2 t=2 2π(ω + αy2t−1 )
Ta Quoc Bao (IU) FRM2 47 / 56
and, the log likelihood
T
T −1 1 X y2t 
log(L(ω, α)) = − log(2π) − log(ω + αy2t−1 ) +
2 2
t=2
ω + αy2t−1

Application to GARCH(1,1)

σt2 = ω + αy2t−1 + βσt−1

2

the density

1  1 y22 
f (y2 | y1 ) = q exp −
2π(ω + αy21 + β σ̂12 ) 2 ω + αy21 + β σ̂12

Ta Quoc Bao (IU) FRM2 48 / 56

and the log likelihood

T −1
log(L(ω, α)) = − log(2π)
2
T
1 X y2t 
− log(ω + αy2t−1 + β σ̂t−1
2
)+
2
t=2
2
ω + αy2t−1 + β σ̂t−1

The importance of σ1
• σ1 can make a large difference
• Especially when the sample size is small
• Typically set σ1 = σ̂

Ta Quoc Bao (IU) FRM2 49 / 56

Volatility targeting
• Since we have the long-run variance rate
ω
σ2 =
1−α−β
• we can set
ω = σ̂ 2 (1 − α − β)
where σ̂ 2 is is the sample variance
• Hence we save one parameter in the estimation

Ta Quoc Bao (IU) FRM2 50 / 56

2.5. Forecasting future volatility
The variance rate estimated at the end of day n − 1 for n day when
apply GARCH(1,1) model is

σn2 = ω + αy2n−1 + βσn−1

2
= σ 2 (1 − α − β) + αy2n−1 + βσn−1
2

or
σn2 − σ 2 = α(y2n−1 − σ 2 ) + β(σn−1
2
− σ2)
On day n + t in the future we have
2
σn+t − σ 2 = α(y2n+t−1 − σ 2 ) + β(σn+t−1
2
− σ2)

Hence,
2
E[σn+t − σ 2 ] = (α + β)E[σn+t−1
2
− σ2]

Ta Quoc Bao (IU) FRM2 51 / 56

By induction we obtain
2
E(σn+t ) = σ 2 + (α + β)t (σn2 − σ 2 ) (2)

Example
For the S&P data consider earlier, α + β = 0.9935, the long-run
variance rate σ 2 = 0.0002075 (or σ = 1.44% per day). Suppose that our
estimate of the current variance rate per day is 0.0003 (This
corresponds to a volatility of 1.732% per day). In t = 10 days, calculate
the expected variance rate??
We have σn2 = 0.0003
Hence
2
E(σn+10 ) = 0.0002075 + 0.993510 × (0.0003 − 0.0002075) = 0.0002942
√
or the expected volatility per day is 0.0002942 = 1.72%, still above the
long-term volatility of 1.44% per day.
Ta Quoc Bao (IU) FRM2 52 / 56
Volatility term structures
Suppose it is day n. We define
2
V(t) = E(σn+1 )

and  1 
a := log
α+β
From (2) we have

V(t) = σ 2 + e−at (V(0) − σ 2 )

Then we have the average variance rate per day between today and
time T

Ta Quoc Bao (IU) FRM2 53 / 56

 Z T Z T
1 1  
V(t)dt = σ 2 + e−at (V(0) − σ 2 ) dt
T 0 T 0
1 − e−aT
= σ2 + [V(0) − σ 2 ]
aT
Now we define σ(T) the volatility per annum that should be used to
price a T-day option under GARCH(1,1) model. Then we have
 1 − e−aT 
σ 2 (T) = 252 σ 2 + [V(0) − σ 2 ] (3)
aT
This relationship between the volatility of options and their maturities is
referred to as the volatility term structure.

Ta Quoc Bao (IU) FRM2 54 / 56

Example
For S&P data, using GARCH(1,1) model we obtain the coefficients
ω = 0.0000013465, α = 0.083394 and β = b = 0.910116. So from (3),
assume that V(0) = 0.0003 we have

0.0000013465
σ2 = = 0.0002073
1 − 0.083394 − 0.910116
and a = log(1/0.99351) = 0.00651. Hence,
 1 − e−0.00651×T 
σ 2 (T) = 252 0.0002073 + [0.0003 − 0.0002073]
0.00651 × T
For the option life (days) T = 10, 30, 50, 100, 500, we obtain the option
volatility (% per annum)
Option life (days) 10 30 50 100 500
option volatility 27.36 27.10 26.87 26.35 24.32

Ta Quoc Bao (IU) FRM2 55 / 56

Calculate the skewness of the distribution

3
f (x) = x2
8
for 0 < x < 2 and f (x) = 0 otherwise

Ta Quoc Bao (IU) FRM2 56 / 56