Volatility

Modeling
Generalized Auto
Regressive Conditional
Heteroscedasticity
(GARCH) Models
 Daily SENSEX
SENSEX (Daily)
20000
18000 Non-stationary Series
16000 Residual Series of ARIMA
14000
12000 1,200
(1,0,0) SARIMA(0,1,1)5:
10000
0.1
estimated by OLS.
8000
Sensex Returns (daily)
800
6000
4000
0.05
Low volatility
2000
400 This suggests residuals are not
0 0 constant over time. In4-Sep-08
fact, it 4-Nov-08
0 4-Mar-08 4-May-08 4-Jul-08
3-Mar-08 3-May-08
is highly3-Sep-08
3-Jul-08
volatile over the
3-Nov-08

-0.05
-400 time. But volatility pattern is
different at different
Stationary but points in
-800-0.1 time.
Volatile Series
-1,200 High volatility
-0.15
I II III IV
 Daily Oil Price
International Crude Oil Price
160
Non-stationary Series
Residual series of
140
20
120
100
16
ARIMA(1,0,1) SARIMA(1,1,1)5
80
60 12
0.2
40
8
High volatility
20

Oil Price Changes period

0.15
0
3-Mar-08 4 0.1
3-May-08 3-Jul-08 3-Sep-08 3-Nov-08

0 0.05

0
This suggests residuals are highly
-4 4-Mar-08 volatile
4-May-08 towards
4-Jul-08 the end, but less4-Nov-08
4-Sep-08
-0.05
-8 volatile in the middle and
-0.1
Stationary
beginning of Series
the sample.
-12
-0.15
I II III IV
2008
 Daily Exchange Rate
Exchage Rate (Daily)
60
Non-stationary Series

Residual series of
50
1.2
40
ARIMA(1,0,0) SARIMA(0,1,1)
Exchange Rate Changes (Daily) 5
30 0.8 3.0%

20 2.0%
0.4
10
1.0%
0.0
0
0.0%
3-Mar-08 3-May-08 3-Jul-08 3-Sep-08 3-Nov-08
4-Mar-08 4-May-08 4-Jul-08 4-Sep-08 4-Nov-08
-0.4
-1.0%

-0.8 -2.0%
This suggests residuals are volatile
-1.2
-3.0%
in the middle and HighofVolatile
end the Period
-4.0%
-1.6
I
sample.II But volatility towards
III
the IV
end is higher than2008the middle.
D(EX,0,5) Residuals
 Monthly Peak Gold Price
2000
Peak-month
1800

1600
Peak-month
Non-stationary Series
1400

1200

1000 250
800
200 Residual series of ARIMA(1,0,1)
600
150 0.2 SARIMA(1,1,1)12
400
100 0.15 Seems…there
Monthly is
Gold Price Changes
200

0 50
0.1 variability, but
October-…

46 October-…
Septemb…
February…

21 Septemb…
26 February…
Decemb…

36 Decemb…
Novemb…

11 Novemb…
variability is constant
April-04

16 April-09
July-05

31 July-10
June-03

6 June-08
March-07

51 March-12
August-07

56 August-12
January-03

1January-08

61January-13
May-06

41 May-11
0.05
0
0
over the sample?
This suggests variability in

66
71
76
81
86
91
96
101
106
111
116
121
-50
-0.05
-100 -0.1 residuals is constant over the
-150 -0.15
sample period unlike ARIMA
2005 models of ‘SENSEX’, 2009OIL2010
PRICE’
-200 -0.2
2006 2007 2008 2011 2012 2013
-0.25
and EXCHANGE RATE’ series.
D(PEAK,0,12) Residuals
 Problem with ARIMA Modeling
• SENSEX returns and Oil price changes/ returns are
stationary but volatile series.
• Univariate ARIMA modelling to forecast the daily
‘SENSEX’ OR ‘Oil Price’ using OLS estimation
will predict the ‘conditional mean’ of these
series based on past information with the
assumption that the residual variance is constant
over time.
• But these series exhibit high variability
suggesting that conclusion/forecast based
on OLS is likely to be inaccurate, hence a
correction is required to reduce this
variability or volatility or risk or
heteroscedastic error variance .
 Need for Volatility Modeling
• The basic version of least squares model assumes
that the expected value of all error terms, when
squared, is same at any given point. The variances of
error term is constant. Known as “HOMOSCEDASTIC”
assumption.

• This assumption often gets violated in energy and

financial time series data where the variance of error
terms are larger at some points than for others. It
exhibits periods of high volatility followed by low
volatility ----implying that the data suffers from
heteroscedaticity.
Homoskedastic Case
Heteroskedastic Case
 Need for Volatility Modeling

• For high frequency time series data, large and

small errors tend to occur in clusters, i.e.,
large returns are followed by more large
returns, and small returns by more small
returns in financial markets.

• Hence, some time periods are riskier than

others; that is, the expected value of the
magnitude of error terms at some periods is
greater than at others.
 Need for Volatility Modeling
• Moreover, these risky times are not scattered
randomly across data.

• Instead, there is a degree of

autocorrelation in the riskiness of
financial returns.

• Here, volatility/variability/risk is autocorrelated/

systematic!
• Instead of correcting this problem and re-estimating
the conditional mean, the ARCH and GARCH models
address the issue of heteroscedasticity by modelling
the heteroscedastic variance which changes over time.
 Volatility Models

• The goal of such models is to provide a volatility

measure—like a standard deviation—that can be
used in financial decisions concerning risk analysis,
portfolio selection and derivative pricing.

• GARCH family models deal with such non-

linear time-series which model the mean
(Expected Return) and Risk (Variance)
conditional on the past information & then
forecast
 Volatility Models
• Let the dependent variable be labeled rt, the return on
an asset or portfolio. The mean value mt and the
variance ht will be defined relative to a past information
set.

• So,
rt = E (rt / rt-1, rt-2, … rt-s) + SE (rt / rt-1, rt-2, …, rt-s) * ut ;
= mt + ht1/2 ut ;

where ut is the error term for the present period,

hence white noise with mean zero and constant variance
 Volatility Models
• The econometric challenge is to specify how
the information is used to forecast the
mean and variance of the return,
conditional on the past information
• While many specifications have been
considered for the ‘mean return’ and have
been used in efforts to forecast future
returns, virtually no methods were
available to model and forecast ‘variance
return’ before the introduction of ARCH
models
• The primary descriptive tool was the
rolling standard deviation. This is the
standard deviation calculated using a fixed
 Volatility Models
• For example, this could be calculated every day
using the most recent month (22 business days)
of data
• It is convenient to think of this formulation as the
first ARCH model; it assumes that the variance
of tomorrow’s return is an equally weighted
average of that for last 22 days
• The assumption of equal weights seems
unattractive, as one would think that the
more recent events would be more relevant
and therefore should have higher weights
• Furthermore the assumption of zero
weights for observations more than one
month old is also unattractive.
 ARCH Models
• ARCH model proposed by Engle (1982) let these
weights be ‘parameters’ of changing variance
to be estimated.

• The model allowed the data to determine the

best weights to use in forecasting the variance.

• ARCH models are employed commonly in

modeling financial time series that exhibit time-
varying volatility clustering,
clustering i.e. periods of
swings followed by periods of relative calm.
 ARCH Models
Consider a regression or auto-regression model;
yt = α + βxt + ut ; OR yt = βyt-1 + ut ;

where ut ~ N(0, σt2) and σt2 is not constant but changes

over time and dependent on the past history.
2
So, ut  t ht ; ht  t

• where εt is white noise and follows standard normal

with mean zero and variance unity and
• ht is the systematic variance which changes over time
 ARCH Models
Engel (1982) proposed a model where volatility in the
current period can be written as follows

ht = σt2 = γ0 + γ1 u2t-1 ; is called an ARCH (1) process.

So, when a big shock happens in previous period (t-1), it is

more likely that the value of ut-1 (because of squares)
will be bigger as well and hence the shock will be large
in current period also if γ1 is positive.
 ARCH Models
• Similarly, ARCH(2) process can be expressed as
ht = γ0 + γ1 u2t-1 + γ2 u2t-2

• ARCH(q) process will be

ht = γ0 + γ1 u2t-1 + γ2 u2t-2 …..+ γq u2t-q + w ;

where wt is a new white noise process with mean zero

and constant variance.

As ht cannot be negative, the ARCH(q) model valid

provided γ0 > 0 and γj >= 0 for j = 1, 2, …, q . Also, for
stationarity conditions, ∑ γ < 1; j = 1, 2, …, q
 Estimation Steps

• Step 1: The underlying variable/s should be stationary

• Step 2 : Estimate the regression or auto-regression
model with OLS and obtain its residuals.
• Step 3: Test for the presence of heteroscedasticity in
residual variance known as “Test of ARCH Effect”
• Step 4: If the ARCH effect is present, estimate an ARCH
model of appropriate lag (q) based on AIC/SBC criteria.
• ARCH model is estimated with Maximum Likelihood
Estimator.
 Estimation Steps
AR(q)-ARCH(q) Model
Estimate the best fitting AR(q) model . So, the
assumption is …process needs to be stationary.

Obtain the squares of the error εt 2 and regress them on a

Constant and q lagged values: where q is the length of
ARCH lags.

under the null hypothesis that conditional

Heteroscedasticity or ARCH disturbance effect is present.
 Estimation Steps
(yt = βxt + ut)- AR(1)-ARCH (q) Model
Estimate the model by OLS (yt = βxt + εt) with one
autoregressive lagged term.
Obtain the squared error term εt 2 and regress εt 2 on a
constant and q lagged values: where q is the length of
ARCH lags.

under the null hypothesis that conditional

heteroscedasticity or ARCH disturbance effect is present.
is present.
 ARCH Effect Test
Obtain the squared error term εt 2 (obtained from
regression or auto-regression model) and regress εt 2 on a
constant and q lagged values: where q is the length of
ARCH lags.

Test there is no autocorrelation in error variance, that is, H0

: α1 = α2 = …= αq = 0 against
H1 : α1 ≠ α2 ≠ … ≠ αq ≠ 0,
This can be tested by both F-statistic and Chi Square
Statistic = nR2 . If accepted, then Var (εt 2 ) = α0 implying
that there is no ARCH effect.
 Important Notes

• In the absence of ARCH components, ARCH (q) = AR(q)

• ARCH model is more of a moving average specification
than a autoregression!!
• Do not confuse with autocorrelation of the error term
with ARCH model. In ARCH model, it is the variance of
error term that depends on the previous error term.
• The presence of ARCH effect does not invalidate the
OLS estimates, but leads to invalid hypothesis testing.
 Problems of ARCH Models
• The intuition behind the ARCH(1) model is that
short-run conditional variance/volatility of the
series is a function of the immediate past values
of the squared error term

• ARCH (q) is an extension of ARCH(1) model, and

useful when volatility of the series is expected to
change more slowly than in ARCH(1) model.
Works best when volatility is a short-memory
process!!
• In case of a long memory process, ARCH(q)
models are difficult to estimate!!
 GARCH Models
• A useful generalization of ARCH model is the
GARCH parameterization introduced by
Bollerslev (1986)

• It gives parsimonious models that are easy to

estimate and, even in its simplest form, has
proven surprisingly successful in predicting
conditional variances
 GARCH Models
• The most widely used GARCH specification asserts that
the best predictor of the variance in the next period is
a weighted average of the long-run average variance,
the variance predicted for this period, and the new
information in this period that is captured by the most
recent squared residual

• Consider the trader who knows that the long-run

average daily standard deviation of the Standard and
Poor’s 500 is 1 percent that the forecast he made
yesterday was 2 percent and the unexpected return
observed today is 3 percent.
 GARCH Models

• Obviously, this is a high volatility period, and

today is especially volatile, which suggests
that the forecast for tomorrow could be even
higher

• However, the fact that the long-term average

is only 1 percent might lead the forecaster to
lower the forecast

• The best strategy depends upon the

dependence between days
 GARCH Models
• If these three numbers are each squared and
weighted equally, then the new forecast would
be (1  4  9) / 3 2.16

• However, rather than weighting these equally,

it is generally found for daily data that weights
such as those in the empirical example of
(.02, .9, .08) are much more accurate

• Hence the forecast is

(0.02 *1  0.9 * 4  0.08 * 9) 2.08
 GARCH (p,q) Models
• This model GARCH (p, q) estimates conditional
variance as a function of ‘weighted average of the
past squared residuals’ till q lagged term, and ‘lagged
conditional variance’ till p terms.
• The basic ARCH (q) model is a short memory process
in that only the most recent q squared residuals are
used to estimate the changing variance.

• The GARCH model allows long memory

processes, which use all the past squared
residuals to estimate the current variance.
 GARCH (p,q) Models
Consider a regression or auto-regression model;
yt = α + βxt + ut ; OR yt = βyt-1 + ut ;
where ut ~ N(0, σ2) and σ2 is not constant but changes over
time and dependent on the past history.
2
ut  t ht ; ht  t

where εt is white noise and ~N(0,1) ;

ht is the systematic variance which changes over
time, a scaling factor.
The GARCH (1,1) model can be written
2 as
ht 0  0ut  1  1ht  1
 GARCH (p,q) Models
The GARCH (p,q) model can be written as
p q
ht 0    i ht  i    j u 2
t j
i 1 j 1

So, ht now depends both on past values of the

shocks/error, which are captured by the lagged squared
residual terms, and on past values of itself, which is
captured by lagged ht terms. This is called Variance
Equation.
γ0 >0, δi > 0, γj > 0; non-negativity conditions.
and δi + γj < 1
Mean Equation is :
yt = α + βxt + ut ; OR yt = βyt-1 + ut ;
 GARCH (p,q) Models
The GARCH (1,1) model can be expressed as an infinite
ARCH (q) process.
GARCH(1,1) Model:
ht 0  1ut2 1  1ht  1 ;
GARCH (1,1) model is equivalent to an infinite order ARCH model
0  1ut2 1  1 (0  1ut2 2  1ht  2 )
with geometrically declining coefficients. This means GARCH (1,1)
modelisan
0   u 2

alternative
1 t 1  
to
1 0   
higher u 2
order
1 1 t 2   2
ARCH
1 ( 0   u
model,2
 1ht less
1 t  2where 3)
parameters are required to be 2estimated.2
0  10  1 0  1ut  1  11ut  2  1 1ut  2
2 2 2

 13 (0  1ut2 4  1ht  4 )

..........
0 
 1  1 ut  j ;
j 1 2

1  1 j 1
 GARCH (p,q) Models
The GARCH (p,q) model
p q
ht 0    i ht  i    j u 2
t j
i 1 j 1
can be expressed as an ARMA (m,p) process in ut2
where m = max{p,q}
p as follows:
p p
ut2 0    j ut2 j    i ut2 i    i wt  i  wt ;
j 1 i 1 i 1
m p
 ut2 0    j ut2 j  wt   w i t i ;
j 1 i 1

where
wt ut2  ht  t2 ht  ht ( t2  1)ht ;  t ~ N (0,1)

The above expression is meaningful but difficult to deal

 GARCH-M (p,q) Models
Yt = α + βXt + θht + ut ; where ut ~ iid N(0,ht)
p q
ht 0    i ht  i    j ut2 j
i 1 j 1

• GARCH in mean or GARCH-M models are those

where conditional variance is a regressor in
conditional mean equation

• Risk-averse investors requires a premium to hold a

risky asset. Risky assets should have higher premium,
higher the risk , higher the premium. So, return is a
function of risk.
• GARCH-M models can be linked with CAPM models
 Problem of ARCH/GARCH
• A major restriction of ARCH/GARCH model is the fact
that these models provide symmetric estimates of
volatility/shocks as residuals are squared term.

• But, in practice, financial markets react differently in

response to “Good News” and “Bad News”.

• Negative shocks (“Bad News”) bring larger impact on

volatility in returns/financial series compared to positive
shocks (“Good News”) having same magnitude

• For many stocks, there is a strong negative correlation

between the current return and the future volatility.
 Problem of ARCH/GARCH
• The tendency for volatility to decline when returns rise
and to rise when returns fall is often known as ‘Leverage
Effect’.

• These issues are handled by T-GARCH (Threshold) and

E-GARCH (Exponential) models
 EGARCH (p,q) Models
Exponential GARCH or EGARCH model was developed by
Nelson (1991) can be written as
q
ut  q
ut  p
log(ht ) 0    j  j    i log(ht  i )
j j

j 1 h t j j 1 h t j i 1

Since the left hand side of the equation is log of ht , so, the
variance itself will be positive irrespective of whether the
coefficients are positive or not.

So, as opposed to the GARCH model, no restrictions

of non-negativity need to be imposed on for the
estimation.
 EGARCH (p,q) Models
EGARCH (p,q) models can be written as
q
ut  q
ut  p
log(ht ) 0    j  j    i log(ht  i )
j j

j 1 h t j j 1 h t j i 1

γ0 : the mean of the volatility equation

λ : represents the size effect, which indicates how much volatility
increases irrespective of the direction of the shock
γ : represents the sign effect, which examine whether shocks have
asymmetric or symmetric effects on volatility.
δ : represents an evaluation of the persistence of shocks. Nelson
shows that absolute value of ‌ δ ‌<1 ensures stationarity.
 EGARCH (p,q) Models
EGARCH (p,q) models can be written as

q
ut  q
ut  p
log(ht ) 0    j  j    i log(ht  i )
j j

j 1 h t j j 1 h t j i 1

A number of researchers have found that evidence of

asymmetry in stock price behaviour-negative surprises seem
to increase volatility more than positive surprises.

Since a lower stock price reduces the value of equity relative to

corporate debt, a sharp decline in stock prices increases
corporate leverage and could thus increase the risk of holding
stocks. So, γ < 0 is sometimes described as ‘leverage effect’
 TGARCH (p,q) Models
The threshold GARCH or TGARCH model was introduced
by Zakoian (1990) and Glosten, Jaganathan and Runkle
(1993)
A TGARCH (1,1) is given by :
ht 0  1ut2 1  1ut2 1d t  1  1ht  1 ;

where dt-1 = 1; if ut-1 < 0 Negative Shock

= 0; if ut-1 >= 0 Positive Shock

is a multiplicative dummy variable to check whether

there is statistically significant difference when shocks
 TGARCH (p,q) Models
• So, when ut-1 >=0 then the effect of ut-1 on ht is γ1 .
When ut-1 < 0, then the effect of ut-1 on ht is (γ1 + λ1 ).
• If λ1 > 0, we conclude that there is asymmetry, otherwise
when λ1 = 0, the news impact is symmetric.
• So, impact of ‘good-news’ and ‘bad-news’ is different. ‘Bad-
news’ has larger effect on volatility of the series than the good
news.

A TGARCH (p,q) model can be written as :

q p
ht 0   ( j   j d t  j )ut2 j    i ht  i ;
j 1 i 1
 I-GARCH (p, q) Models
• In financial time series, the conditional volatility is
persistent. For a long time series model, if one
estimates GARCH (1,1) model, it will be found that the
sum of γ1 + δ1 is very close to unity.

• This constraint forces the conditional variance to act

like a process with unit root.

• One of the limitations of GARCH (p,q) models is it

assumes that the process is stationary. I-GARCH or
Integrated GARCH overcomes this.
 I-GARCH (p, q) Models

• A GARCH (p,q) process is stationary with a finite

variance if p q


i1
i  j 1
j1

• A GARCH (p, q) process is I-GARCH or non-stationary

GARCH (p,q) if

p q


i1
i    j 1
j1
 I-GARCH (p, q) Models
• I-GARCH processes are either non-stationary or if they
are stationary they have infinite variance. Infinite
variance means heavy tails! A distribution can be heavy-
tailed with a finite variance.

• A GARCH (1,1) model can be expressed as :

0 
 1  1 ut  j ;
j 1 2
ht 
1  1 j 1

so, unlike a true non-stationary process, the conditional

variance ht is a geometrically decaying function of the
current and past realizations of ut-j2 . Hence, an I-GARCH
model can be estimated like any other GARCH model.