Time Series Econometrics

ARCH Models

Walter Sosa Escudero

Universidad de San Andr¶
es and UNLP

Motivation
-2 0 2 4 6
-6

0 50 100 150 200

-2 0 2 4 6
-6

0 50 100 150 200

² Which are the main di®erences/similarities between these

two processes?

Walter Sosa Escudero. Topics in Applied Econometrics 2

Close up: points 45 to 110

5
0
es[45:110]

-5
-10

50 60 70 80 90 100 110

45:110

² This is an example of an ARCH(1) process.

² Does the variance of this process change? What about the
conditional variance?

Walter Sosa Escudero. Topics in Applied Econometrics 3

Motivation: conditional and unconditional models

Consider the simple stationary AR(1) model:

Yt = ÁYt¡1 + "t ; jÁj < 1
Its unconditional mean and variance are:
²E(Yt ) = 0
²V (Yt ) = ¾2 =(1 ¡ Á2 )
Its conditional mean and variance are:
²Et (Yt ) = ÁYt¡1
²Vt (Yt ) = ¾2
Et stands for the expectation conditional on the information
available at period t (past information).

Walter Sosa Escudero. Topics in Applied Econometrics 4

 ² This simple speci¯cation has a varying conditional
mean and a constant conditional variance. A feature of
many economic series (in°ation, stock returns, etc.) is
that its conditional variance varies over time: periods of
large volatility tend to be followed by periods of large
volatility.
² In many cases it is relevant to forecast the conditional
volatility of prices. It is consistent with forming rational
expectations.
² Modeling the conditional variance correctly might
improve the precision of estimates or forecasts of
conditional means.
² Another interesting feature of many ¯nancial series is
that they have `fat tails'. This might come from
varying conditional variances.

Walter Sosa Escudero. Topics in Applied Econometrics 5

Aside: The law of iterated expectations

Let Y and X be two random variables. Then:

E(Y ) = E[E(Y jX)]

This is the law of iterated expectations.

² It allows to compute unconditional expectations in

`two-steps': ¯rst compute the conditional expectation
(which is a random variable), and then the expectation
of this conditional expectation.

Walter Sosa Escudero. Topics in Applied Econometrics 6

 The Basic ARCH speci¯cation

Consider the following model:

q
Yt = "t 2 ; " » W N (0; 1)
®0 + ®1 Yt¡1 t

and let us compute the conditional moments:

² Et (Yt ) = 0
² Vt (Yt ) = Et (Yt2 ) ´ h(t) = ®0 + ®1 Yt¡1
2

The unconditional moments are:

² E(Yt ) = E(Et (Yt )) = 0
² V (Yt ) = E(Yt2 ) = E(Et (Yt2 )) = ®0 + ®1 E(Yt¡1
2
)
®0
If j®1 j < 1,, V (Yt ) = E(Yt2 ) = 1¡®1

Walter Sosa Escudero. Topics in Applied Econometrics 7

q
The process Yt = "t 2 , known as pure ARCH(1)
®0 + ®1 Yt¡1
has:

² Constant and zero unconditional mean and constant

unconditional variance.

² Constant zero conditional mean and varying conditional

variance

² The conditional variance depend on how large where

2
deviations from the mean (Yt¡1 ) in the previous periods.

² If ®1 = 0 this is a simple white noise process.

Walter Sosa Escudero. Topics in Applied Econometrics 8

 The ARCH(q) regression model

Yt = mt + "t ; "t j−t » (0; ht )

ht = ®0 + ®1 "2t¡1 + ¢ ¢ ¢ + ®q "2t¡q
−t refers to all the information available at period t (past
information)
Examples:

² mt = 0 is our previous `pure ARCH' case.

² mt = x0t ¯ , this is a simple linear regression model with

ARCH(q) residuals.

² mt = ÁYt¡1 + µ"t¡1 , this is an ARMA(1,1) model with

ARCH(q) residuals.

Walter Sosa Escudero. Topics in Applied Econometrics 9

ARCH and non-normality

Consider the basic ARCH(1) model:

Yt j−t » N (0; ht ); ht = ®0 + ®1 "2t
were we have assumed normality. Remember that if X is a
normal random variable then:
² Skewness = E(Yt ¡ ¹)3 =¾ 3 = 0.
² Kurtosis = E(Yt ¡ ¹)4 =¾ 4 = 3.

For the ARCH case:

² E(Yt3 ) = E(Et (Yt3 )) = 0
² E(Yt4 )=¾ 4 = 3(1 ¡ ®21 )=(1 ¡ 3®12 ) > 3
The ARCH process has non-normal unconditional
distribution. It has thicker tails than the normal.

Walter Sosa Escudero. Topics in Applied Econometrics 10

 ML estimation of ARCH models

Consider the ARCH regression model under conditional

normality:

yt j−t » N (x0t ¯; ht ); ht = ®0 +®1 "2t¡1 +¢ ¢ ¢+®q "2t¡q ; "t ´ yt ¡x0t ¯

The sample log-likelihood conditional on the ¯rst q

observations will be:

T T
T 1X 1 X (yt ¡ x0t ¯)2
l(µ) = ¡ ln(2¼) ¡ ln ht ¡
2 2 t=1 2 t=1 ht

which can be maximized using standard numerical methods.

Walter Sosa Escudero. Topics in Applied Econometrics 11

Testing for ARCH e®ects

A simple test for ARCH e®ects can be based on the

Lagrange Multiplier principle under conditional normality:

1. Estimate the regression model by OLS and save

residuals et .

2. Regress et on e2t¡1 ; e2t¡2 ; : : : e2t¡q and keep the R2 .

3. Under the null of no ARCH e®ects, T R2 » Â2 (q)

asymptotically.

Walter Sosa Escudero. Topics in Applied Econometrics 12

 GARCH models

In practice volatility tends to be very persistent. It requires

many lags in the ARCH speci¯cation. In addition, it
requires many restrictions to have a positive variance.
Bollerslev (1986), proposes the GARCH(p,q) speci¯cation:
q
X p
X
h t = ®0 + ®i "2t¡i + ¯i ht¡i
i=1 i=1

with ®0 > 0; ®i ; ¯i ¸ 0
² The conditional variance depends on past squared shocks
and on past values of the conditional variance too.
² As in the ARMA case, it allows for a more parsimonious
representation of ARCH models: `small' GARCH
speci¯cations work better than `long' ARCH.

Walter Sosa Escudero. Topics in Applied Econometrics 13

Let vt ´ "2t ¡ ht . Replace ht = "2t ¡ vt in the GARCH

speci¯cation:

q
X p
X
"2t = ®0 + ®i "2t¡i + ¯i ("2t¡i ¡ vt¡i ) + vt
i=1 i=1

Then, the GARCH(p,q) is actually an ARMA(m,p) for "2t ,

with m = max(p; q).

² Be careful, the ARMA speci¯cation refers to "2t and not

to ht .

² In particular, for p = 0 it implies that the ARCH(q)

speci¯cation is actually an AR(q) speci¯cation for "2t

Walter Sosa Escudero. Topics in Applied Econometrics 14

 The ARCH-M model

Engle, Lilien and Robins (1987): the mean depends on the

conditional variance too.
Yt = x0t ¯ + ±ht + "t
p
"t = ht vt
ht = ®0 + ®1 "2t¡1 + ¢ ¢ ¢ + ®q "2t¡q
Here the conditional variance a®ects the conditional mean
through the ±ht term. A simple test for ARCH-M e®ects
can be based on testing the null ± = 0 in the full estimation
of the model.

Walter Sosa Escudero. Topics in Applied Econometrics 15

Varieties of ARCH speci¯cations

² IGARCH (integrated GARCH): consider the GARCH

speci¯cation:
q
X p
X
ht = ®0 + ®i "2t¡i + ¯i ht¡i
i=1 i=1

The IGARCH corresponds to setting

Pq Pp
i=1 ®i + i=1 ¯i = 1. This has proved useful to model
high dependences in volatility.

² TARCH (threshold ARCH): consider the process:

ht = ®0 + ®1 "t¡1 + ¸dt¡1 "2t¡1 + ¯1 ht¡1

where dt¡1 = 1 if "t¡1 < 0 and 0 otherwise.

Walter Sosa Escudero. Topics in Applied Econometrics 16

² TARCH (continued). This allows for a di®erent e®ect
on the conditional variance if errors are positive or
negative. Positive ¸ means that negative shocks have
larger e®ects on volatility than positive shocks. A test
of asymmetric e®ects checks ¸ = 0.

² EGARCH (exponential GARCH).

¯ ¯
"t¡1 ¯ " ¯
¯ t¡1 ¯
ln ht = ®0 + ®1 p + ¸ ¯p ¯ + ¯1 ln ht¡1
ht¡1 ¯ ht¡1 ¯

Coe±cients are not constrained to be positive. It uses

standardized lags of the shocks. It allows for
asymmetric e®ects.

Walter Sosa Escudero. Topics in Applied Econometrics 17