ES1004 Econometrics by Example

Lecture 11: ARCH and GARCH Models

Dr. Hany Abdel-Latif

Swansea University, UK

Gujarati textbook, second edition [chapter 15]

15th October 2016

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 1 / 38

Time Series Econometrics

13 stationary and nonstationary time series

14 cointegration and error correction models
15 asset price volatility: the ARCH and GARCH models
16 economic forecasting

previous course on time series econometrics

ES1002 Lectures

ES1002 EViews

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 2 / 38

 Financial Time Series Properties

Volatility Clustering

financial time series, such as stock prices, interest rates, foreign

exchange rates, often exhibit volatility clustering
periods of turbulence: prices show wide swings; and
periods of tranquillity: there is a relative calm

various sources of news and other economic events may have an

impact on the time series pattern of asset prices
news can lead to various interpretations, and economic events like an
oil crisis can last for some time
so we often observe the large positive and large negative observations
in financial time series to appear in clusters

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 3 / 38

 Financial Time Series Properties

Real and Financial Impacts

such swings in oil prices and credit crises have serious effects
investors are concerned about the
rate of return on their investment
risk of investment and the variability or volatility of risk
it is important to measure asset price and asset returns volatility

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 4 / 38

 Financial Time Series Properties

Measuring Volatility

a simple measure of asset return volatility is its variance over time

variance by itself does not capture volatility clustering
subtract the mean value from individual values, square the difference
and divide it by the number of observations
a measure of unconditional variance
a single number of a given sample
does not take into account the past history (time-varying volatility)

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 5 / 38

 ARCH Meaning

The ARCH Model

autoregressive conditional heteroscedasticity

a measure that takes into account the past history (time-varying
volatility)
in time series data involving asset returns, such as returns on stocks
or foreign exchange, we observe autocorrelated heteroscedasticity

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 6 / 38

 ARCH Meaning

Autocorrelated Heteroscedasticity

heteroscedasticity, or unequal variance, in cross section data because

of the heterogeneity among individual cross-section units
in time series data, we usually observe autocorrelation
in financial data we observe autocorrelated heteroscedasticity
i.e., heteroscedasticity observed over different periods is autocorrelated
in the literature, this phenomenon called ARCH effect

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 7 / 38

 ARCH Example

Example: Exchange Rate

data table13_1.xls
the exchange rate between the us dollar and the euro EX; dollars per
unit of euro
daily from January 4, 2000 to May 8,2008 [2355 observations]
are not continuous; exchange rate markets are not always open every
day and because of holidays
see figure next slide

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 8 / 38

 ARCH Example

Exchange Rate: Log

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 9 / 38

 ARCH Example

Exchange Rate: Log-changes

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 10 / 38

 ARCH Volatility

Variance vs. Volatility

the variance of a random variable is a measure of the variability in the

values of the random variable
for our data on daily exchange rate returns
the mean is about 0.000113 or 0.0113%
the variance is about 0.0000351
this variance does not capture the volatility of the daily exchange rate
return seen in previous figure
because it does not take into account the variation in the amplitudes
noticed in the figure

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 11 / 38

 ARCH Volatility

Measuring Volatility

a simple way to measure volatility

RETt = c + ut

where RTEt daily return, c a constant, ut error term

if we obtain the residuals et and square them, you get the plot in the
next slide

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 12 / 38

 ARCH Volatility

Regression

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 13 / 38

 ARCH Volatility

Squared Residuals

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 14 / 38

 ARCH Volatility

Measuring Volatility

wide swings in the squared residuals can be taken as an indicator for

underlying volatility
in the squared residual figure observe there
clusters of periods when volatility is high and clusters of periods when
volatility is low
these clusters seems to be autocorrelated
when volatility is high, it continues to be high for quite some time
when volatility is low, it continues to be low for a while

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 15 / 38

 ARCH The Model

The ARCH Model

Yt |It−1 = α + βXt + ut

Yt exchange rate return, Xt one variable or a vector of variables

conditional on the information available up to time (t − 1), the value

of the random variable Yt is a function of the variable Xt and ut

ut |It−1 ∼ iid N(0, σt2 )

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 16 / 38

 ARCH The Model

The ARCH Model

in the CLARM it is assumed that σt2 = σ 2 homoscedastic variance

but to take into account the ARCH effect, we let

σt2 = λ0 + λ1 ut−1
2

we assume that the error variance at time t is equal to some constant

plus a constant multiplied by the squared error term in the previous
time period

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 17 / 38

 ARCH The Model

The ARCH Model

σt2 = λ0 + λ1 ut−1
2

if λ1 = 0
- the error variance is homoscedastic
- the framework of the CLRM applies

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 18 / 38

 ARCH The Model

The ARCH Model

σt2 = λ0 + λ1 ut−1
2

coefficients of this equation should be positive because the variance

cannot be a negative number
it is assumed that 0 < λ1 < 1

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 19 / 38

 ARCH The Model

The ARCH Model

Yt |It−1 = α + βXt + ut

after taking the mathematical expectation on both sides

α + βXt the conditional mean equation
σt2 = λ0 + 2
λ1 ut−1 the conditional variance equation
both conditional on the information set It−1

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 20 / 38

 ARCH The Model

ARCH(p)

σt2 = λ0 + λ1 ut−1
2

this equation known as ARCH(1) model

includes only one lagged squared value of the error term
this model can be easily extended to an ARCH(p) model, where we
have p lagged squared error terms

σt2 = λ0 + λ1 ut−1
2 2
+ λ2 ut−2 2
+ · · · + λp ut−p

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 21 / 38

 ARCH The Model

Testing ARCH Effect

σt2 = λ0 + λ1 ut−1
2 2
+ λ2 ut−2 2
+ · · · + λp ut−p

if there is an ARCH effect, it can be tested by the statistical

significance of the estimated λ coefficients
if they are significantly different from zero, we can conclude that
there is an ARCH effect

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 22 / 38

 ARCH The Model

Estimation

σt2 = λ0 + λ1 ut−1
2 2
+ λ2 ut−2 2
+ · · · + λp ut−p

since the u are not directly observable, we use the estimated residuals

û = Yt − α̂t − β X̂t

then we estimate the following model

ût2 = λ0 + λ1 ût−1
2 2
+ λ2 ût−2 2
+ · · · + λp ût−p

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 23 / 38

 ARCH The Model

The ARCH Model

ût2 = λ0 + λ1 ût−1
2 2
+ λ2 ût−2 2
+ · · · + λp ût−p

AR we are regressing squared residuals on its lagged values going

back to p periods
CH variance is conditional on the information available up to time
(t − 1)

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 24 / 38

 ARCH The Model

ARCH(8) OLS Estimation

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 25 / 38

 ARCH The Model

Estimation of ARCH Model

the maximum likelihood approach

an advantage of the ML method is that we estimate the mean and
variance functions simultaneously
statistical packages such as stata and eviews, have built-in routines to
estimate ARCH models

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 26 / 38

 ARCH The Model

ARCH(8) ML Estimation

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 27 / 38

 ARCH The Model

ARCH(8) ML Estimation

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 28 / 38

 ARCH The Model

ARCH(8) ML Estimation

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 29 / 38

 ARCH The Model

ARCH(8) ML Estimation

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 30 / 38

 ARCH Drawbacks

Drawbacks of ARCH Model

requires estimation of the coefficients of p autoregressive terms, which

consumes several degrees of freedom
difficult to interpret all the coefficients, especially if some of them are
negative
the OLS estimating procedure does not lend itself to estimate the
mean and variance function simultaneously
the literature suggests that any model higher than ARCH(3) is better
estimated by GARCH

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 31 / 38

 GARCH The Model

GARCH Model

generalised autoregressive conditional heteroscedasticity

we modify the variance equation to get GARCH(1,1) as follows

σt2 = λ0 + λ1 ut−1
2 2
+ λ2 σt−1

conditional variance at time t depends on

the lagged squared error term at time (t − 1), and
the lagged variance term at time (t − 1)

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 32 / 38

 GARCH The Model

GARCH(1,1)

σt2 = λ0 + λ1 ut−1
2 2
+ λ2 σt−1

it can be shown that ARCH(p) model is equivalent to GARCH(1,1) as

p increases
in ARCH(p) we have to estimate (p + 1) coefficients, whereas in
GARCH(1,1) model we estimate only 3 coefficients
GARCH(1,1) can be extended to GARCH(p,q) model
p lagged squared error terms
q lagged conditional variance terms
in practice, GARCH(1,1) has proved useful to model returns on
financial assets

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 33 / 38

 GARCH Example

GARCH(1,1)

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 34 / 38

 GARCH Example

GARCH(1,1)

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 35 / 38

 GARCH Extensions GARCH-M

The GARCH-M Model

modify the mean equation by explicitly introducing the risk factor, the
conditional variance, to take into account the risk

Yt = α + βXt + γσt2 + ut

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 36 / 38

 GARCH Extensions GARCH-M

GARCH-M(1,1) Example

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 37 / 38

 GARCH Extensions GARCH-M

GARCH-M(1,1) Example

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 38 / 38

 Questions & Answers

Dr. Hany Abdel-Latif (2016) ES1004ebe Lecture 11 ARCH & GARCH 39 / 38