You are on page 1of 9

GARCH Modelling

Introduction
It was first observed by Engle (1982) that although many financial time series, such as, stock returns and exchange rates are unpredictable, there is apparent clustering in the variability or volatility. This is often referred to as conditional heteroscedascity since it is assumed that overall the series is stationary but the the conditional expected value of the variance may be time-dependent. Many different time series models have been suggested. We will only study the ARCH/GARCH family. The ARCH/GARCH is not available in most standard statistical software systems such as R, S-Plus, SPSS or Mathematica. But it is available in the S-Plus module Finmetrics and is the free software JMulti. RATS is another software package with ARCH/GARCH models. There are many other more specialized packages for sale.

ARCHHqL Model
Engle (1982) first suggested the ARCH model. at = xt st , xt ~NIDH0, 1L

where z > 0 and ai 0, i = 1, , q. The conditions z > 0 and ai 0, i = 1, , q are sufficient (but not necessary) to ensure that st > 0. The ARCH model is always stationary. Often xt is assumed t-distributed or some other symmetric fat tailed distribution.

s2 = z + a1 a2 + + aq a2 t t-1 t-q Hvolatility equationL

Unconditional Mean, Variance and Kurtosis


Mean < at > = < xt at > = << xt st >t-1 > = << x >t-1 st > = < 0 st > = 0 Variance s2 = < a2 > = < < a2 >t-1 > = < w + a1 a2 + + aq a2 > a t t t-1 t-q = z + a1 < a2 > + + aq < a2 > t-1 t-q = z + Ha1 + + aq L s2 a

Hence,

z s2 = . a 1 - a1 - - aq

ARCH.nb

So in order for s2 > 0, a1 + + aq < 1. a Remark Kurtosis Additional restrictions on a's are required to ensure higher-order moments exist. Taking xt ~ NIDH0, 1L and assuming for simplicity q = 1 and that m4 = < u4 > exists, t m4 = < a4 > t = << a4 >t-1 > t = << xt4 s4 >t-1 > t

Solving for m4 ,

General::spell1 : Possible spelling error: new symbol name "zeta" is similar to existing symbol "Zeta".

2 alpha1 y i z + 3 alpha12 m4, m4E z j SolveAm4 == 3 zeta2 j1 + 1 alpha1 { k 99m4 99m4 3 H1 + alpha1L zeta2 == H1 + alpha1L H1 + 3 alpha12 L

= 3 < Hz + a1 a2 L > t-1 = 3 < z 2 + 2 a1 z a2 + a2 a4 > t-1 1 t-1 2 ga1 y 2i j = 3 z j1 + z + 3 a2 m4 z 1 1 - ga1 { k


2

More

% . 8m4 m4 , zeta , alpha1 1 < 3 2 H1 + 1 L == H1 + 1 L H1 + 3 2 L 1

So

Hence for the fourth moment to exist we need, 0 a2 < 1 3. 1 The standardized kurtosis, k = m4 m2 - 3, 2
3 + 6 alpha12 1 3 alpha12 H1 + alpha1L H1 + 3 alpha1 L
2

3 z 2 H1 + a1 L m4 = H1 - a1 L H1 - 3 a2 L 1

3 H1 + alpha1L zeta2

Hzeta H1 alpha1LL2 FullSimplify

So k > 0, which implies that ut has thicker tails than the normal distribution Hk = 0L. This fits that fact that long financial time series, usually have thick tails and outliers.

ARCH.nb

AR Representation
Setting vt = a2 - s2 then vt is white noise, that is, < vt > = 0 and < vt , vs > = 0, t s. t t Sub-ing s2 = a2 - vt , obtain, t t a2 = z + a1 a2 + + aq a2 + vt t t-1 t-q or in operator notation, aHBL = 1 - a1 B - - aq Bq and aHBL a2 = z + vt . t Hence we can summary that if at ~ARCHHqL then a2 ~ARHqL. t So we see that a2 exhibits mean-reversion, that is, < a2 > s2 as l . t a t+l Further consistent and Gaussian-efficient estimates for z, a1 , , aq can be obtained by the usual AR estimators. Remark: A necessary condition for a2 to be stationary is that all roots of aHBL = 0 are outside the unit circle. But this t always holds when ai 0, i = 1, , q and a1 + + aq < 1.

Estimation
Gaussian LS We can obtain Gaussian efficient estimates by fitting an ARHqL to the squares of the series, a2 t MLE Let t-1 = 8a0 , a1 , , at-1 <. Then the joint density function can be written, Given data a1 , ..., an the loglikelihood L = log f Han-k n-k-1 L
n-1 k=0

f Ha1 , , at L = f Hat t-1 L f Hat-1 t-2 L ... f Ha1 0 L f Ha0 L

Other Distributions

can be evaluated using the model volatility equation for any assumed distribution for the xt . Then L can be Then maximized numerically to obtain the MLE. Let q = Hz, a1 , , aq L denote the parameters. !!! ` n Iq - qM NH0, -1 L, where is the Fisher information matrix. Rather than taking xt ~ NIDH0, 1L, we may use other distributions. t-distribution on v degrees of freedom,
v f HxL I M x2 +v
v+1 2

ARCH.nb

The parameter v may also be estimated by MLE along with the other parameters. QMLE Using the MLE under for the model with xt ~ NIDH0, 1L, produces consistent and asymptotically normal estimates with the same covariance even when xt is IID mean zero and finite variance from some other distribution such as the t. However the same is not true if we use MLE for some other distribution and it is misspecified.

Diagnostic Checking
Test the standardized residuals for autocorrelation, ` ` at = at st

as well as the squared values.

Model Construction
Mean effect model: ARIMA(p,d,q) if there is autocorrelation in the series zt at , where at is white noise Volatility model: ARCH(q) if there is autocorrelation in the squared residuals The order q may be selected by examining a PACF of a2 . t

GARCH Models
Often ARCHHqL models require a large number of parameters. To reduce the number of parameters the GARCHHq, pL was introduced (Bollerslev, 1986). at = xt st , xt ~NIDH0, 1L s2 = z + a1 a2 + + aq a2 + b1 s2 + + b p s2 p t t-1 t-q t-1 tSufficient conditions for positive s2 are z > 0, and ai ' s and bi 's all nonnegative. t As with ARCH, other distributions such as the t-distribution are often used for xt .

Variance
Assuming s2 = < a2 > exists, a t < a2 > = << xt2 s2 >t-1 > t t = << xt2 Hz + a1 a2 + + aq a2 + b1 s2 + + b p s2 p L >t-1 > t-1 t-q t-1 t= z + Ha1 + + aq L s2 + Hb1 + + b p L s2 a a = < Hz + a1 a2 + + aq a2 + b1 s2 + + b p s2 p L > t-1 t-q t-1 t-

ARCH.nb

So z s2 = a 1 - a1 - - aq - b1 - - b p So in order for the variance, s2 , to exist we need a a1 + + aq + b1 + + b p < 1

Kurtosis
For GARCH(1,1),
3 z H1+a1 +b L b -3 a m4 = < a4 > = -2 a12L t H1-a -b L H1-b2
2 1 1 1 1 1 1

So m4 exists only if b2 + 2 a1 b1 + 3 a2 < 1 1 1

And if may be shown that k = m4 m2 - 3 > 0, where m2 = s2 . a 2

Exercise: Derive m4 from first principles and show the k > 3. Peggy

Conditional Variance as Sum of Past Squared Innovations


In operator notation, with bHBL = 1 - b1 B - - b p B p and aHBL = a1 B + + aq Bq So s2 = z + yHBL a2 , where yHBL = aHBL bHBL . Provided bHBL is invertible. So t t s2 = z + yi a2 t t-i
i=1

bHBL s2 = z + aHBL a2 t t

ARMA Representation
Setting vt = a2 - s2 then vt is white noise, that is, < vt > = 0 and < vt , vs > = 0, t s. t t Sub-ing s2 = a2 - vt , in the model equation, obtain, t t a2 = z + a1 a2 + + aq a2 + b1 Ha2 - vt-1 L + + b p Ha2 p - vt- p L t t-1 t-q t-1 ta2 t = z + Hai + bi L a2 + vt - bi vt-i t-i
q p i=1 i=1

So the GARCHHq, pL can be represented as an ARMAHq , pL, where q = maxH p, qL,

where ai = 0 for i > q and bi = 0 for i > p. In other words, at ~GARCHHq, pL then a2 ~ARMAHq , pL. Consist tent estimators of the parameters can be obtained by fitting an ARMAHq , pL to the squared series. However this method is statistically inefficient.

ARCH.nb

GARCH(1,1) Model
The GARCH(1,1) is the most important GARCH model. Higher-order GARCH models tend to be overcomplicated and are rarely used in practice. So in the GARCH(1,1) case, at = xt st , xt ~NIDH0, 1L s2 = z + a1 a2 + b1 s2 t t-1 t-1 where z > 0 and a1 , b1 0. Stationarity In the normal case, a1 + b1 < 1 is sufficient for covariance stationary, ergodic process. In the case of non-normal distributions, the condition is b1 + a1 VarHxt L < 1, where xt are IID with mean zero and finite variance. Under slightly weaker conditions, at may be ergodic and strictly stationary but not covariance stationary. s2 H1L = z + a1 a2 + b1 s2 t t t

Volatility Forecasting

Exercise: Derive from first principles and explain why this is exactly analagous for the forecasts using the ARMA(1,1) model. Ellen.

s2 HlL = z + Ha1 + b1 L s2 Hl - 1L t t

Let s2 HlL denote the MMSE forecast of s2 given all the information up to and including time t. Then, t t+l

IGARCH(1,1) Model
GARCH(1,1) model but with a1 + b1 = 1. This model is not covariance stationary but it is ergodic and strictly stationary. This model seems to occur often in actual applications.

TGARCH
It has been observed empirically and also from finance theory that stock returns volatility is much more effected by negative news than by positive news. This asymmetry in response is called the leverage effect. TGARCH or threshold GARCH was introduced to take into account the asymmetry in response of financial markets to positive and negative news, represented by the innovations, at . s2 = z + a1 a2 + a- Hat-1 < 0L + b1 s2 t t-1 1 t-1

Again z > 0 and all other parameters are nonnegative. The leverage effect is present when a- > 0. 1

Example: Canada/US Exchange Rate

ARCH.nb

<< ARBuild.m
Updated October 16, 2006. Loaded ARBuild

z = ReadData@"d:math2006529adataCanUSexrate.dat"D;
anadaUS exchange rates, daily about 200 missing values dropped Source: JMulti Number of Values Input = 8515

TimeSeriesPlot@z, 104 , 0.7D;

1.6 1.5 1.4 1.3 1.2 1.1 1 0 2000 4000 6000 8000

TimeSeriesPlot@w = TSDifference@zD, 104 , 0.7D;

ARCH.nb

0.04

0.02

-0.02

-0.04

0
SacfPlot@wD;

2000

4000

6000

8000

0 1

10

12

14

0.5

0.5

Acf

-0.5

-0.5

-1 0 2 4 6 8 Lag 10 12 14

SacfPlot@w ^ 2D;

ARCH.nb

0 1

10

12

14

0.5 Acf

0.5

-0.5

-0.5

-1 0 2 4 6 8 Lag 10 12 14

PacfPlot@w ^ 2, SimultaneousBound TrueD;

0 1

10

12

14

0.5 Acf

0.5

-0.5

-0.5

-1 0 2 4 6 8 Lag 10 12 14

You might also like