VOLATILITY MODELS

MEASURING AND FORECASTING VOLATILITY

BEFORE GARCH
Although volatility is viewed as a standard deviation, we
will formulate models for the variance and then take square
roots.
Moving variance: rt is the excess return,
1 N 2
2

r

t N 1
t j
j 0
1) all observations from t-N to t are given equal weight
2) all observations before t-N are given no weight
3) the choice of N is left to the trader.

EXPONENTIAL SMOOTHING

2t 2t 1 (1 ) rt21
Once is known and an initial variance is given such as from
the first few observations, then all the variances can be simply
computed using something like a spreadsheet.
This model is the same as:

1 j 1 rt 2 j
2
t

j 1

a moving variance model with declining weights, no truncation

point, and fixed . This model predicts that all future variances
will be the same as today's variance.

GARCH
GENERALIZED- more general than ARCH
model
AUTOREGRESSIVE-depends on its own
past
CONDITIONAL-variance depends upon
past information
HETEROSKEDASTICITY- fancy word for
non-constant variance

HISTORY
I DEVELOPED THE ARCH MODEL
WHEN I WAS VISITING LSE IN 1979
IT WAS PUBLISHED IN 1982 WITH
MACRO APPLICATION - THE
VARIANCE OF UK INFLATION
TIM BOLLERSLEV DEVELOPED THE
GARCH GENERALIZATION AS MY
PHD STUDENT - PUBLISHED IN 1986

THE FIRST FINANCIAL APPLICATIONS:

Engle Lilien and Robins(1987)
French, Schwert and Stambaugh(1987),
Bollerslev(1988) and Bollerslev, Engle, Wooldridge(1988).
SURVEYS:
Bollerslev, Chou and Kroner(1992), Journal of Econometrics,
Special Issue on ARCH Models in Finance
Bollerslev, Engle and Nelson,(1994), ARCH Models, in
Handbook of Econometrics, volume IV, (eds. Engle and
McFadden), Elsevier.
Engle, ARCH: SELECTED READINGS, Oxford University
Press, 1995

The GARCH Model

rt

ht t

ht ht 1 t21 ht 1
rt21 ht 1

The variance of rt is a weighted average of

three components
a constant or unconditional variance
yesterdays forecast
yesterdays news

This model can be rewritten as:

ht

j 1
(1 )

j 1

rt 2 j

It can also be rewritten:

ht 1 ht ht t2 1

The forecast of conditional variance one step ahead is given

by the first square bracket and the surprise is given by the
second. Volatility is predictable but not perfectly.

PARAMETER ESTIMATION
Historical data reveals when volatilities
were large and the process of volatility
Pick parameters to match the historical
volatility episodes
Maximum Likelihood is a systematic
approach:
rt mt 2
1
Max L( ) log ht

ht

DIAGNOSTIC CHECKING
Time varying volatility is revealed by
volatility clusters
These are measured by the Ljung Box
statistic on squared returns
The standardized returns r / h no longer
t
t
should show significant volatilty clustering
Best models will minimize AIC and
Schwarz criteria

THEOREMS
GARCH MODELS WITH GAUSSIAN
SHOCKS HAVE EXCESS KURTOSIS
FORECASTS OF GARCH(1,1) ARE
MONOTONICALLY INCREASING OR
DECREASING IN HORIZON

FORECASTING WITH GARCH

rt2 )rt21 (rt21 ht 1) (rt2 ht )
GARCH(1,1) can be written as ARMA(1,1)
The autoregressive coefficient is ( )

The moving average coefficient is

GARCH(1,1) Forecasts
ht ( r ) ( ht 1 )
2
t 1

Et ht k Et ht k 1

Monotonic Term Structure of

Volatility

FORECAST PERIOD

FORECASTING AVERAGE
VOLATILITY
Et rt 1 ... rt k Et (rt 1) ... Et (rt k )
2

Et (ht 1) ... Et (ht k )

Annualized Vol=square root of 252 times the
average daily standard deviation
Assume that returns are uncorrelated.

TWO YEARS TERM

STRUCTURE OF PORT
0.14

0.24

0.13

0.22

0.12
0.20
0.11
0.18
0.10
0.16

0.09
0.08
2000 2050 2100 2150 2200 2250 2300 2350 2400 2450 2500
TERM2000

0.14
2900 2950 3000 3050 3100 3150 3200 3250 3300 3350
TERMEND

0.025

0.020

0.015

0.010

0.005
500
0.178

1000

1500

0.22

2000
0.188

0.176

0.186

0.174

0.21

0.184

0.172
0.170

0.182

0.20

0.180

0.168
0.166

0.178

0.19

0.164

0.176

0.162
2450 2500 2550 2600 2650 2700 2750 2800 2850 2900
TERMMIL_2411

0.18
1800 1850 1900 1950 2000 2050 2100 2150 2200 2250 2300
TERMMIL_1800

0.174
2400 2450 2500 2550 2600 2650 2700 2750 2800 2850
TERMMIL_2357

Variance Targeting
Rewriting the GARCH model
ht (ht 1 t21 ) (ht 1 )

where /(1 ) is easily seen to be

the unconditional or long run variance
this parameter can be constrained to be equal
to some number such as the sample variance.
MLE only estimates the dynamics

The Component Model

ht

2
qt ( rt 1 qt 1)

qt

(ht 1 qt 1 )

2
(qt 1 ) (rt 1 ht 1 )

Engle and Lee(1999)

q is long run component and (h-q) is transitory
volatility mean reverts to a slowly moving long
run component

GARCH(p,q)
rt ht t
ht

2
j ht j t j
j 1

j ht j
j 1

The Leverage Effect Asymmetric Models

Engle and Ng(1993) following Nelson(1989)
News Impact Curve relates todays returns to tomorrows
volatility
Define d as a dummy variable which is 1 for down days

ht

2
2
rt 1 rt 1dt 1 ht 1

NEWS IMPACT CURVE

VOLATILITY

NEWS

Other Asymmetric Models

EGARCH: NELSON(1989)
rt 1
rt 1
log(ht ) log(ht 1 )

h t 1
ht 1

NGARCH: ENGLE(1990)
ht (rt 1 ) 2 ht 1

PARTIALLY NON-PARAMETRIC
ENGLE AND NG(1993)
VOLATILITY

NEWS

EXOGENOUS VARIABLES IN
A GARCH MODEL
Include predetermined variables into the
variance equation
Easy to estimate and forecast one step
Multi-step forecasting is difficult
Timing may not be right

ht

2
rt 1

ht 1 zt 1

EXAMPLES
Non-linear effects
Deterministic Effects
News from other markets

Heat waves vs. Meteor Showers

Other assets
Implied Volatilities
Index volatility

MacroVariables or Events

WHAT IS THE BEST MODEL?

The most reliable and robust is
GARCH(1,1)
For short term forecasts, this is good
enough.
For long term forecasts, a component model
with leverage is often needed.
A model with economic causal variables is
the ideal

Component with Leverage

2
2
ht qt (rt 1 qt 1) (ht 1 qt 1) (rt 1dt 1 .5qt 1)
2
qt (qt 1 ) (rt 1 ht 1)

dt 1 if rt 0

Procter and Gamble Daily

Returns 2/89-3/99
Variance Equation Garch(1,1)
C
ARCH(1)

5.11E-06
0.047402

1.24E-06
0.006681

4.112268
7.094764

GARCH(1)

0.927946

0.011114

83.49235

AIC -5.7086 , SCHWARZ CRITERION -5.699617

CORRELOGRAM OF
SQUARED RESIDUALS
AC PAC Q-Stat Prob
1 0.030
0.030 2.3720 0.124
2 -0.018
-0.019 3.1992 0.202
3 -0.010
-0.009 3.4859 0.323
4 0.018
0.018 4.3016 0.367
5 -0.011
-0.013 4.6317 0.462
6 -0.002
-0.001 4.6408 0.591
7 0.013
0.013 5.0865 0.649
8 -0.016
-0.017 5.7164 0.679
9 0.010
0.012 5.9973 0.740
10 0.003
0.002 6.0183 0.814

P&G TARCH
C
0.0000 0.0000 4.6121 0.0000
ARCH(1)
0.0269 0.0093 2.9062 0.0037
(RESID<0)*ARCH(1)0.0520 0.0141 3.6976 0.0002
GARCH(1)
0.9123 0.0139 65.8060 0.0000
AIC -5.7114 SCHWARZ CRITERION -5.7002

P&G EGARCH

C
|RES|/SQR[GARCH](1)
RES/SQR[GARCH](1)

-0.3836
0.1186
-0.0392

0.0672 -5.7052
0.0153 7.7645
0.0103 -3.8195

EGARCH(1)

0.9656

0.0070 137.9063

AIC -5.7114

SCHWARZ CRITERION -5.7002

P&G Asymmetric Component

Perm: C
0.0002 0.0000 14.1047
Perm: [Q-C] 0.9835 0.0049 201.5877
Perm: [ARCH-GARCH]
0.0335 0.0079 4.2577
Tran: [ARCH-Q]
-0.0361 0.018 -2.0045
Tran: (RES<0)*[ARCH-Q] 0.0910 0.0213 4.2838
Tran: [GARCH-Q] 0.8063 0.0819 9.8403
AIC -5.7132,

SCHWARZ CRITERION -5.6974

P&G VOLATILITIES
0.7
0.6
0.5
0.4
0.3
0.2
0.1
2/22/89

1/23/91

V_PG_TARCH

12/23/92 11/23/94 10/23/96

V_PG_GARCH

9/23/98

V_PG_ACOMP

DISCUSS GAUSSIAN
ASSUMPTION
EMPIRICAL EVIDENCE INDICATES THAT
INNOVATIONS ARE LEPTOKURTIC
GAUSSIAN GARCH IS QMLE
CONSISTENT BUT NEEDS BOLLERSLEV
WOOLDRIDGE STANDARD ERRORS
T-DISTRIBUTION MAY BE MORE EFFICIENT
CAN DO SEMI-PARAMETRIC ESTIMATOR OF
ENGLE AND GONZALES-RIVERA

Bollerslev Wooldridge
Standard Errors
ROBUST TO NON-NORMAL ERRORS
Perm: C
Perm: [Q-C]
Perm: [ARCH-GARCH]
Tran: [ARCH-Q]
Tran: (RES<0)*[ARCH-Q]
Tran: [GARCH-Q]

0.0002 0.0000 8.3857

0.9835 0.0081 121.9762
0.0335 0.0107 3.1264
-0.036 0.0242 -1.4898
0.0910 0.0429 2.1220
0.8063 0.1223 6.5919

RISK PREMIA
WHEN RISK IS GREATER, EXPECTED
RETURNS SHOULD BE GREATER
HOW MUCH?
WHAT COUNTS AS RISK?

CAPM GIVES AN ANSWER

MULTI-BETA GIVES ANOTHER
PRICING KERNEL COVERS ALL