Conditional Heteroskedasticity

Time Series Analysis Stylized facts on Volatility Clustering Realized Variance ARCH models GARCH models I-GARCH, EGARCH
Lecture 5: Conditional Heteroskedasticity Models

Drew Creal
February 2, 2016
Drew Creal (Chicago Booth)
February 2, 2016
1 / 104
Outline
1
Time Series Analysis
Stylized facts on Volatility Clustering
Realized Variance
ARCH models
GARCH models
I-GARCH, EGARCH, GARCH-M, GJR models
Value-at-Risk
References
February 2, 2016
2 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
3 / 104
ARMA(p,q)
I
so far, we have focused on modeling the conditional mean:

rt = E [rt |rt 1 , rt 2 , . . . ; ] + t
the benchmark models for the conditional mean are ARMA models
rt = 0 + 1 rt 1 + . . . + p rt p + t 1 t 1 + . . . q t q
given estimates, 0 , 1 , . . . , q 1 , q , we can calculate the residuals

t = vt (0) recursively from the initial condition.
the key modeling goal is to make sure that the residuals { t } are
white noise
February 2, 2016
4 / 104
Why White Noise?

I
consider an AR(1):
rt = 0 + 1 rt 1 + t
suppose the AR(1) is misspecified
then the forecast errors are correlated over time, i.e.:

vt (0) = t
vt +1 (0) = t +1
..
.
that means your forecast is not optimal (youre not using all the
information optimally)
forecast errors should be white noise

February 2, 2016
5 / 104
Volatility Clustering
suppose we have a good model for the conditional mean t (e.g.

ARMA(p, q ))
now we look at the squared residuals {2t } to test for conditional

heteroskedasticity
for returns on most financial assets, there is a lot more

autocorrelation in conditional second moments than in the
conditional first moments.
volatility is predictable!
February 2, 2016
6 / 104
Volatility?
In most asset markets, volatility varies dramatically over time.
episodes of high volatility seem to be clustered.
we dont simply observe volatility
how do we measure vol?

1. implied volatility (back out volatility from option prices)
2. (non-parametric) realized volatility (e.g., realized volatility of stock
returns over one-month using daily data)
3. (parametric) model volatility X
February 2, 2016
7 / 104
VIX-Implied Volatility
VIX CBOE SPX Volatility
70
60
50
40
30
20
10
1987
1990
1993
1995
1998
2001
2004
2006
2009
2012
Annualized measure of 30-day vol. VIX, designed to measure the markets expectation
of 30-day volatility implied by at-the-money S&P 100 Index (OEX) option prices.
Monthly data. 1990-2009. VIX white paper
February 2, 2016
8 / 104
Why do we care about Volatility?
why do we care about modeling vol?

1. portfolio allocation
2. risk management
3. option pricing
February 2, 2016
9 / 104
ARCH model of Engle (1982)

Definition
The ARCH(m) model is given by:
t = t t ,
t2 = 0 + 1 2t 1 + . . . + m 2t m
where t is i.i.d., mean zero, has variance of one , 0 > 0 and i 0 for
i >0
I
the standard normal is a common choice for t
the (standardized) Students t is another option for t
large shocks are followed by more large shocks
February 2, 2016
10 / 104
Building a Volatility Model
1. estimate a model for the conditional mean (e.g. ARMA model)

2. use the residuals t from the mean equation to test for ARCH effects
3. specify a volatility model for ARCH effects
4. check the fitted model
February 2, 2016
11 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
12 / 104
Monthly Returns on Market Squared

log returns on market squared
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
1916
1930
1943
1957
1971
1984
1998
2012
2026
log stock returns on market -squared. Monthly data. 1926-2012.
February 2, 2016
13 / 104
Stock Returns
ACF for log returns
ACF for |log returns|
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0.1
10
15
Month
20
25
0.1
ACF for squared returns
10
15
Month
20
25
PACF for squared returns
0.4
0.4
0.3
0.3
0.2
0.2
0.1
0.1
0
0.1
10
15
Month
20
25
0.1
10
15
Month
20
25
log stock returns on CRSP-VW. Monthly data. 1926-2012.
February 2, 2016
14 / 104
Dynamics of Variance-Monthly Returns

ACF for log returns
0.1
0.3
0.05
0.2
0.1
0.05
0.1
0.1
10
15
Month
20
25

0.3
0.3
0.2
0.2
0.1
0.1
0.1
10
15
Month
20
10
15
Month
20
25
0.4
0.1
25
0.2
10
15
Month
20
25
February 2, 2016
15 / 104
Dynamics of Variance-Monthly Returns

ACF for log returns
0.1
0.25
0.2
0.05
0.15
0.1
0
0.05
0
0.05
0.05
0.1
10
15
Month
20
25
0.1

0.15
0.1
0.1
0.05
0.05
0.05
0.05
10
15
Month
20
10
15
Month
20
25
0.15
0.1
25
0.1
10
15
Month
20
25
February 2, 2016
16 / 104
Daily Returns on Market Squared

log daily returns on market squared
0.04
0.035
0.03
0.025
0.02
0.015
0.01
0.005
0
1916
1930
1943
1957
1971
1984
1998
2012
2026
log stock returns on market -squared. Daily data. 1988-2012.
February 2, 2016
17 / 104
Dynamics of Variance -Daily Returns

ACF for log returns
0.08
0.4
0.06
0.3
0.04
0.2
0.02
0.1
0
0
0.02
0.04
10
15
20
25
0.1
10
Days
20
25
0.3
0.3
0.25
0.25
0.2
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0
0.05
15
Days
0
0
10
15
Days
20
25
0.05
10
15
20
25
Days
log stock returns on CRSP-VW. Daily data. 1926-2012.
February 2, 2016
18 / 104
Dynamics of Variance-Daily Returns

ACF for log returns
0.06
0.4
0.04
0.3
0.02
0.2
0.1
0.02
0.04
10
15
20
25
0.1
10
Days
20
25
0.3
0.3
0.25
0.25
0.2
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0
0.05
15
Days
0
0
10
15
Days
20
25
0.05
10
15
20
25
Days
log stock returns on CRSP-VW. Daily data. 1970-2012.
February 2, 2016
19 / 104
IBM
log returns on IBM squared
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
1987
1990
1993
1995
1998
2001
2004
2006
2009
2012
log stock returns on IBM -squared. Monthly data. 1988-2009.
February 2, 2016
20 / 104
IBM
ACF for log returns
0.2
0.2
0.1
0.1
0.1
0.1
0.2
0.2
0
10
15
Month
20
25
ACF for squared log returns
10
15
Month
20
25
PACF for squared log returns
0.2
0.2
0.1
0.1
0.1
0.1
0.2
0.2
0
10
15
Month
20
25
10
15
Month
20
25
log stock returns on IBM. Monthly data. 1988-2009.
February 2, 2016
21 / 104
MSFT
log returns on MSFT squared
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
1987
1990
1993
1995
1998
2001
2004
2006
2009
2012
log stock returns on MSFT -squared. Monthly data. 1988-2012.
February 2, 2016
22 / 104
MSFT
ACF for log returns
0.2
0.2
0.1
0.1
0.1
0.1
0.2
0.2
0
10
15
Month
20
25
ACF for squared log returns
10
15
Month
20
25
PACF for squared log returns
0.2
0.2
0.1
0.1
0.1
0.1
0.2
0.2
0
10
15
Month
20
25
10
15
Month
20
25
log stock returns on MSFT. Monthly data. 1988-2009.
February 2, 2016
23 / 104
Changes in log U.S. $

Changes in log USD (TradeWeighted Index)
0.1
0.08
0.06
0.04
0.02
0.02
0.04
0.06
0.08
0.1
1971
1976
1982
1987
1993
1998
2004
2009
log changes in Dollar -Trade-Weighted Index. Monthly data. 1971-2009.
February 2, 2016
24 / 104
Changes in log U.S. $-squared

Changes in log USD 2
x 10
0
1971
1976
1982
1987
1993
1998
2004
2009
February 2, 2016
25 / 104
U.S. $
ACF for log changes
ACF for |log changes|
0.1
0.15
0.1
0.05
0.05
0
0
0.05
0.1
0.05
10
15
Month
20
25
0.1
ACF for squared log changes

0.15
10
15
Month
20
25
0.15
0.1
0.1
0.05
0.05
0.05
0.05
0.1
PACF for squared log changes
10
15
Month
20
25
0.1
10
15
Month
20
25
February 2, 2016
26 / 104
Yen
log changes in Yen/USD squared
0.01
0.009
0.008
0.007
0.006
0.005
0.004
0.003
0.002
0.001
0
1971
1976
1982
1987
1993
1998
2004
2009
log changes in Yen/USD squared. Daily data. 1971-2009.
February 2, 2016
27 / 104
Yen
ACF for log changes
ACF for |log changes|
0.15
0.3
0.1
0.2
0.05
0.1
0.05
0.1
10
15
20
25
10
15
20
Days
PACF for squared log changes
0.2
0.2
0.15
0.15
0.1
0.1
0.05
0.05
0
0.05
Days
ACF for squared log changes
25
10
15
Days
20
25
0.05
10
15
20
25
Days
log changes in Yen/USD. Daily data. 1971-2009.
February 2, 2016
28 / 104
Testing for ARCH
estimate a model for the conditional mean: e.g. ARMA(p, q )
use the squared residuals to test for ARCH
specify a volatility model if ARCH effects are documented
February 2, 2016
29 / 104
Testing for ARCH(1)

I
we can test for autoregressive conditional heteroscedasticity using a

LB Q-test:
suppose the model for the conditional mean is ARMA(p, q )
estimate the model
run a Q-test on the estimated squared residuals {2t }

I
test the null that:

H0 : 1 = 2 = . . . = m = 0
where i denotes the autocorrelation of the squared residuals

if you cannot reject the null, no need for ARCH machinery
February 2, 2016
30 / 104
Testing for ARCH(2)

I
we can test for autoregressive conditional heteroscedasticity using a

LM test:
suppose the model is ARMA(p, q )
estimate the model

run a LM-test on the estimated squared residuals {2t }
I test the null that = 0, i = 1, 2, . . . , m:
i
2t = 0 + 1 2t 1 + . . . + m 2t m + et , t = m + 1, . . . , T
I
test the null that:

H0 : 1 = 2 = . . . = m = 0
use the usual F-test, reject the null if F exceeds 2 () the (1 )-th
upper percentile
February 2, 2016
31 / 104
Volatility
volatility clusters
volatility is stationary
leverage effect (asymmetry)

I
negative returns seem to be followed by larger increases in volatility

than positive returns
first emphasized by Black (1976)
Black (1976) suggested that negative returns (decreases in price)

change a companys debt/equity ratio, increasing their leverage.
February 2, 2016
32 / 104
Leverage Effects: Monthly Returns on Market Squared

log returns on market squared
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
1916
1930
1943
1957
1971
1984
1998
2012
2026
1998
2012
2026
negative log returns on market

0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
1916
1930
1943
1957
1971
1984
log stock returns on market -squared. Monthly data. 1926-2012.

February 2, 2016
33 / 104
Models of Volatility
Definition
The model is given by
rt = t + t
where t is governed by an ARMA(p, q ) and where the conditional
variance of t varies according to:
V[ t |Ft 1 ] = t2
February 2, 2016
34 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
35 / 104
High-Frequency Data
I let rtm denote the log of the returns over a period of time
I may denote 1 day, 1 week, 1 month.
I suppose over the time , we observe equally spaced log-returns {rt,i }n at a
i =1
higher frequency
rtm =
rt,i
i =1
I for log returns, the variance of the return over time is given by:
V (rtm | Ft 1 ) =
V (rt,i | Ft 1 ) + 2 Cov
i =1
rt,i , rt,j |Ft 1
i <j
I if we assume that the returns at the higher frequency are i.i.d., then
V (rtm |Ft 1 ) = nV (rt,1 |Ft 1 )

where V (rt,1 |Ft 1 ) can be estimated from:
2
b
t,m
=
2
ni=1 (rt,i r t )
n1
February 2, 2016
36 / 104
Realized Variance (Volatility)

I
the estimate for the variance is given by:

2
b
t,m
=
2
ni=1 (rt,i r t )
n1
in practice, is often 1 day or 1 month.
assumption: estimator assumes errors are i.i.d. meaning that volatility

is roughly constant over .
note: the term realized variance is not used consistently in the

literature.
many finance papers will call this realized volatility. In econometrics,

realized volatility is the square root of realized variance.
February 2, 2016
37 / 104
Realized Variance
Realized Vol
100
90
80
Annualized Vol
70
60
50
40
30
20
10
0
1916
1930
1943
1957
1971
1984
1998
2012
2026
Annualized (Monthly) Realized Vol. Daily data. 1926-2012.

February 2, 2016
38 / 104
Realized Variance and Implied Vol
Realized Vol and Implied Vol (VIX)

90
Realized
VIX
80
70
Annualized Vol
60
50
40
30
20
10
0
1982
1987
1993
1998
2004
2009
2015
Annualized (Monthly) Realized Vol and the VIX. Daily data. 1983-2012.
February 2, 2016
39 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
40 / 104
ARCH(1)
Definition
The ARCH(1) process is given by:
t = t t ,
t2 = 0 + 1 2t 1
where t is i.i.d., mean zero, has variance of one , > 0 and i 0 for
i >0
February 2, 2016
41 / 104
Variance
I
the unconditional mean is zero:

E[ t ] = E(E[ t |Ft 1 ]) = E(t E[t |Ft 1 ])
the unconditional variance is:

V[ t ] = E 2t = E E 2t |Ft 1 = E 0 + 1 2t 1
because of stationarity, we get that:

V [ t ] = 0 + 1 V [ t ]
as a result, the variance is:

V[ t ] =
0
1 1
we require that 0 1 < 1

February 2, 2016
42 / 104
Fourth Moment
I
the unconditional fourth moment is :

E[4t ] = E(E[4t |Ft 1 ]) = E(3t4 ) = 3E 0 + 1 2t 1
using stationarity, this implies

m4 =
2
320 (1 + 1 )
(1 1 )(1 321 )
which means 0 21 < 1/3

the unconditional kurtosis is given by:
E[4t ]
1 21
=
3
>3
V [ t ]2
1 321
positive kurtosis even though innovations are conditionally Gaussian

February 2, 2016
43 / 104
Normality
ARCH(1) process is conditionally normal if t N (0, 1)X
ARCH(1) process is not unconditionally normal X

I
time variation in the variance generates fat tails
February 2, 2016
44 / 104
AR format
the ARCH(m ) model in AR format is given by:

2t +1 = 0 + 1 2t + . . . + m 2t m+1 + t +1
where t = 2t t2 is an uncorrelated series with mean zero.
this is like an AR(m ), except that t is not i.i.d.
the PACF is a useful too for determing the right lag length
February 2, 2016
45 / 104
1-step ahead forecast
the ARCH(m ) model is given by:

t2+1 = 0 + 1 2t + . . . + m 2t m+1
hence the forecast is simply:

t2 (1) = 0 + 1 2t + . . . + m 2t m+1
February 2, 2016
46 / 104
2-step ahead forecast

t2+2 = 0 + 1 2t +1 + . . . + m 2t m+2
hence the forecast is simply:

t2 (2) = 0 + 1 t2 (1) + . . . + m 2t m+2
note that we have used the following:

Et 2t +1 |Ft = Et t2+1 t2+1 |Ft = t2+1 Et t2+1 |Ft
February 2, 2016
47 / 104
h-step ahead forecast

t2+h = 0 + 1 2t +h + . . . + m 2t m+h
General expression:
t2 (h ) = 0 + 1 t2 (h 1) + . . . + m t2 (h m )
where t2 (h j ) = 2t +hj if h j < 0
February 2, 2016
48 / 104
Building an ARCH model
we can pick the order of an ARCH model by looking at PACF at

{2t }; you might need a large number of lags
for a well-specified ARCH model, the standardized residual:

e
t =
t
t
should be white noise

I
this can be tested by examining e

t
1. check the volatility spec by doing a Q-test on e

2t
2. check the mean spec by doing a Q-test on e
t
February 2, 2016
49 / 104
Weakness of ARCH Models
1. symmetry in effects of positive and negative shocks on vol

2. ARCH model is restrictive (21 < 1/3 for ARCH(1))
3. ARCH model: mechanical description of volatility (no economics)
4. sometimes many lags are needed to describe vol dynamics X
Matlab function
February 2, 2016
50 / 104
ARCH(5)
Table: ARCH(5)
Parameter
Value
s.e.
t-stat
Constant
ARCH1
ARCH2
ARCH3
ARCH4
ARCH5
3.09E-05
0.099307
0.144208
0.141901
0.168116
0.150002
6.65E-07
0.005071
0.008707
0.009809
0.01012
0.010459
46.472
19.5838
16.5623
14.4668
16.6119
14.3424
ARCH(5). Daily data. CRSP-VW.1971-2012.

February 2, 2016
51 / 104
Parametric Volatility t
Conditional Annualized Volatility
140
arch(5)
120
100
80
60
40
20
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
ARCH(5). Daily data. CRSP-VW.1971-2012. use the [V , logL] = infer (EstMdl, y

command.
February 2, 2016
52 / 104
Parametric Volatility against Realized Vol.

140
realized
arch(5)
120
100
80
60
40
20
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
53 / 104
Standardized Residuals ( t /t )2 .
Standardized Residuals squared
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
54 / 104
Standardized Residuals t /t .
ACF for standardized residuals
ACF for |stand. residuals|
0.15
0.06
0.04
0.1
0.02
0.05
0
0
0.05
0.02
10
15
20
25
0.04
10
15
20
Days
PACF for squared standard residuals
0.04
0.02
0.02
0.02
0.02
ACF for squared stand. residuals

0.04
0.04
Days
10
15
Days
20
25
0.04
10
15
20
25
25
Days
February 2, 2016
55 / 104
AIC/BIC.
check the AIC/BIC:
[V , logL] = infer (EstMdl, y )

[aic, bic ] = aicbic (logL, 6, size (y , 1))
aic = 7.0162e + 04
bic = 7.0126e + 04
February 2, 2016
56 / 104
ARCH(10)
Table: ARCH(10)
Parameter
Value
standard error
t-stat
Constant
ARCH1
ARCH2
ARCH3
ARCH4
ARCH5
ARCH6
ARCH7
ARCH8
ARCH9
ARCH10
2.04E-05
0.083715
0.113967
0.080502
0.095266
0.107659
0.075962
0.052205
0.083885
0.073778
0.049606
7.56E-07
0.004496
0.008698
0.009147
0.00937
0.011581
0.009292
0.008305
0.00877
0.008694
0.008884
26.927
18.6207
13.1022
8.80109
10.1667
9.29659
8.1747
6.28631
9.56551
8.4864
5.58371

February 2, 2016
57 / 104
120
arch(10)
100
80
60
40
20
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
ARCH(10). Daily data. CRSP-VW.1971-2012. use the [V , logL] = infer (EstMdl, y

command.
February 2, 2016
58 / 104
Parametric Volatility against Realized Variance.

120
realized
arch(10)
100
80
60
40
20
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
59 / 104
100
90
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
60 / 104
0.12
0.04
0.1
0.02
0.08
0.06
0
0.04
0.02
0.02
0
0.02
10
15
20
25
0.04
10
15
20
Days
0.02
0.01
0.01
0.01
0.01

0.02
0.02
Days
10
15
Days
20
25
0.02
10
15
20
25
25
Days
February 2, 2016
61 / 104
AIC/BIC.
check the AIC/BIC:

aic = 7.0538e + 04
bic = 7.0465e + 04
I
ARCH(10) has smaller AIC and BIC than ARCH(5)
February 2, 2016
62 / 104
AIC/BIC.
4
6.75
AIC
x 10
6.75
6.8
6.8
6.85
6.85
6.9
6.9
6.95
6.95
7.05
7.05
7.1
10
15
lags
20
25
7.1
BIC
x 10
10
15
20
25
lags
ARCH(lags). Daily data. CRSP-VW.1971-2012.
February 2, 2016
63 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
64 / 104
GARCH model of Bollerslev (1986)
Definition
Consider t = rt t . The t follows a GARCH(m, s ) model if
m
t = t t ,
t2 = 0 + i 2t i +
i =1
j t2j
j =1
where t is i.i.d., mean zero, has variance of one , 0 > 0 and i 0 for
i > 0 and j 0 for j > 0, and
max (m,s )
( i + i ) < 1
i =1
February 2, 2016
65 / 104
GARCH is ARMA for vol

I
define:
t = 2t t2
hence
t2 = 2t t
this delivers an ARMA-like representation:

max (m,s )
2t = 0 +
i =1
(i + i )2t i j t j + t
j =1
where E(t ) = 0 and E(t t j ) = 0 but is not white noise

the unconditional variance is simply:

0
E 2t =
max (m,s )
1 i =1
( i + i )
February 2, 2016
66 / 104
GARCH(1,1) of Bollerslev (1986)

I
consider GARCH(1,1) model:

t
2
t +1
I
= t t
t N (0, 1)
2
= 0 + 1 t + 1 t2
large 2t , t2 leads to large t2+1
the unconditional kurtosis is given by:

E[4t ]
1 ( 1 + 1 )2
=
3
>3
V [ t ]2
1 (1 + 1 )2 221
positive kurtosis even though innovations t are Gaussian
February 2, 2016
67 / 104
Forecasting with GARCH(1,1)

I

t2+1 = 0 + 1 2t + 1 t2
the one-step ahead forecast of volatility:

t2 (1) = 0 + 1 2t + 1 t2
the two-step ahead forecast of volatility:

t2 (2) = 0 + (1 + 1 )[t2 (1)]
where we used the following expression
t2+1 = 0 + (1 + 1 )t2 + 1 t2 2t 1
February 2, 2016
68 / 104
Forecasting with GARCH(1,1)

I

t2+1 = 0 + 1 2t + 1 t2
the one-step ahead forecast of volatility:

t2 (1) = 0 + 1 2t + 1 t2
the two-step ahead forecast of volatility:

t2 (2) = 0 + (1 + 1 )[t2 (1)]
in general:
t2 (h ) = 0 + (1 + 1 )[t2 (h 1)]
February 2, 2016
69 / 104
Forecasting with GARCH(1,1) at longer horizons
Result
The multi-step volatility forecast t2 (h )
t2 (h ) = 0 + (1 + 1 )[t2 (h 1)]
converges to the unconditional variance as h increases
February 2, 2016
70 / 104
Picking Order and Checking Model
for a well-specified GARCH model, the standardized residual:

e
t =
t
t
should be white noise

I
this can be tested by examining e

t
1. check the volatility specification by doing a Q-test on e

2t
2. check the mean spec. by doing a Q-test on e
t
February 2, 2016
71 / 104
GARCH(1,1)
Table: GARCH(1,1)
Parameter
Value
standard error
t-stat
Constant
GARCH1
ARCH1
1.16E-06
0.909294
0.07998
2.23E-07
0.002906
0.001725
5.18121
312.933
46.3601
GARCH(1,1). Daily data. CRSP-VW.1971-2012.
February 2, 2016
72 / 104
90
garch(1,1)
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
GARCH(1,1). Daily data. CRSP-VW.1971-2012. use the [V , logL] = infer (EstMdl, y

command.
February 2, 2016
73 / 104

100
realized
garch(1,1)
90
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
74 / 104
120
100
80
60
40
20
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
75 / 104
0.12
0.03
0.1
0.02
0.08
0.01
0.06
0.04
0.01
0.02
0.02
0
0.02
0.03
0
10
15
20
25
0.04
10
15
20
Days
0.02
0.01
0.01
0.01
0.01

0.02
0.02
Days
10
15
Days
20
25
0.02
10
15
20
25
25
Days
February 2, 2016
76 / 104
AIC/BIC.
check the AIC/BIC:

aic = 7.0738e + 04
bic = 7.0665e + 04
I
GARCH(1,1) has smaller AIC and BIC than ARCH(10)
you could use the BIC/AIC criterion to find the best GARCH(p,q)
February 2, 2016
77 / 104
Forecasting
forecasting using the GARCH(1,1) model
suppose we want to forecast the daily variance over the next 500
periods
use the observed returns y to initialize the forecast

Vf = forecast (EstMdl, 500,0 Y 00 , y )
February 2, 2016
78 / 104
Forecast.
3
x 10
3.5
Conditional Variance Forecasts for VWCRSP Returns

Inferred
Forecast
2.5
1.5
0.5
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
11000
February 2, 2016
79 / 104
Forecast.
5
11
Conditional Variance Forecasts for VWCRSP Returns
x 10
10.5
10
9.5
8.5
7.5
6.5
50
100
150
200
250
300
350
400
450
500
February 2, 2016
80 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
81 / 104
I-GARCH
Definition
Consider t = rt t . The t follows a I -GARCH(1, 1) model if
t = t t ,
t2 = 0 + (1 1 )2t 1 + 1 t21
where t is i.i.d., mean zero, has variance of one. 0 < < 1

I
the unconditional variance is not defined!
hard to justify for returns
February 2, 2016
82 / 104
Forecasting with I-GARCH(1,1)
in general for GARCH(1,1):

t2 (h ) = 0 + (1 + 1 )[t2 (h 1)]
now set 1 + 1 = 1
repeated substitution yields:

t2 (h ) = (h 1)0 + [t2 (1)]
I
effects on future vols are persistent!
February 2, 2016
83 / 104
GARCH(1,1)-M
Definition
Consider
t = rt t t2 .
The t follows a GARCH(1, 1)-M model if
t = t t ,
t2 = 0 + 1 2t 1 + 1 t21
where t is i.i.d., mean zero, has variance of one. 0 < < 1.

M stands for GARCH in the mean.
I
the parameter measures a variance risk premium
the risk premium increases when volatility increases

February 2, 2016
84 / 104
E-GARCH model of Nelson (1991)

Definition
Consider t = rt t . The t follows a E-GARCH(1, 1) model if
t = t t ,
(1 1 B ) ln t2 = (1 )0 + g ( t 1 )
where t is i.i.d., mean zero, has variance of one.

g ( t 1 ) = ( + ) t E(| t |) if t 0
g ( t 1 ) = ( )et E(| t |) if t < 0
I
built-in asymmetry to capture leverage effect
we expect < 0
February 2, 2016
85 / 104
E-GARCH(1,1)
Table: E-GARCH(1,1)
Parameter
Value
standard error
t-stat
Constant
GARCH1
ARCH1
Leverage1
-0.17256
0.980926
0.13937
-0.07601
0.011062
0.001135
0.005217
0.003161
-15.599
864.546
26.7121
-24.0473
EGARCH(1,1). Daily data. CRSP-VW.1971-2012.
February 2, 2016
86 / 104
90
egarch(1,1)
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
EGARCH(1,1). Daily data. CRSP-VW.1971-2012. use the [V , logL] = infer (EstMdl, y

command.
February 2, 2016
87 / 104

90
realized
egarch(1,1)
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
88 / 104
100
90
80
70
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
February 2, 2016
89 / 104
0.12
0.04
0.1
0.02
0.08
0.06
0.04
0.02
0.02
0.04
0
0.02
10
15
20
25
0.06
10
15
20
Days
0.02
0.01
0.01
0.01
0.01

0.02
0.02
Days
10
15
Days
20
25
0.02
10
15
20
25
25
Days
February 2, 2016
90 / 104
Glosten, Jagannathan, and Runkle (1993)

Definition
Consider t = rt t . The t follows a GJR-GARCH(1, 1) model if
t = t t ,
t2 = 0 + 1 2t 1 + 1 t21 + It 1
where t is i.i.d., mean zero, has variance of one.

0
if t 0
It =
1
if t < 0
I
simple way to capture leverage effect
we expect > 0
February 2, 2016
91 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
92 / 104
VaR
Definition
Let the r.v. Lt,t +h denote the loss of a financial position over a time
horizon t t + h. The Value at Risk over the time horizon h is the
amount of money you could lose with probability p. This is given by:
p = Pr [Lt,t +h VaR ] = 1 Pr [Lt,t +h < VaR ] .
I For a given probability p, how much could a position lose? This is the amount of
money (value) that is at risk.

I in applications, we need
1.
2.
3.
4.
5.
the
the
the
the
the
probability (say p = .01)

time horizon h (say h = 10 days)
frequency of the data
loss cdf Ft ,t +h
size of the position
February 2, 2016
93 / 104
Remarks
By convention, VaR is reported as a positive number.
VaR is the 1 p quantile function of the loss distribution.
For a long position, a loss is a negative return.
Option #1: we can look at the distribution of rt and calculate the

1 p-th quantile.
Option #2: we can look at the distribution of rt , calculate the p-th

quantile, and take the absolute value.
Warning: be careful when calculating this. The distribution of returns

is typically not symmetric.
February 2, 2016
94 / 104
Simple example
I
Suppose you bought $10000 of MSFT.
Let rt +1 denote the log-return over the next month. Our model is:
rt +1 N (0.05, 0.01)
The one-period return CDF is: Ft,t +1 (rt +1 ) = (rt +1 |0.05, 0.01)
1
Let p = 0.01. Calculate the quantile Ft,t
+1 (0.01) = 0.1145. The
value at risk is:
VaR0.01 = |10000 0.1145| = 1145

I
1
Let p = 0.05. Calculate the quantile Ft,t
+1 (0.05) = 0.1826. The
value at risk is:
VaR0.05 = |10000 0.1826| = 1826

February 2, 2016
95 / 104
Comments
The quantile function is known (or easily calculated) when returns are
conditionally normal or conditionally Students t.
In most models used for risk management, the conditional distribution

of returns is not known. We need to calculate VaR by simulation
methods.
RiskMetrics uses the IGARCH(1,1) model with normally distributed

errors t and no drift t = 0. This is a special case where returns are
conditionally normal at all horizons. (Next example.)
February 2, 2016
96 / 104
RiskMetrics
Definition
The RiskMetrics Value at Risk assumes the daily log return
rt |Ft 1 N (t , t2 )
where
t = 0,
t2 = t21 + (1 )rt21 , 1 > > 0
Hence,
rt = t = t t
t N (0, 1)
is an IGARCH(1,1) process without drift.

I
The combination of no drift, normal errors, and IGARCH implies that

the conditional distribution is known at all horizons.
February 2, 2016
97 / 104
VaR in RiskMetrics
I
Using the independence assumption for t :

h
V (rt [h ]|Ft ) =
V ( t +i |Ft )
i =1

where V ( t +i |Ft ) = E t2+i |Ft can be computed recursively.
from the I-GARCH(1,1) specification:
t2 = t21 + (1 )t21 (t21 1) t
this implies that the conditional variance can be stated as:
t2+i = t2+i 1 + (1 )t2+i 1 (t2+i 1 1),

i = 2, . . . , h

and hence E t2+i |Ft = t2+1 , i = 2, . . . , h
hence, the conditional variance of the long-horizon return is given by:
V (rt [h ]| Ft ) = hV ( t +1 | Ft ) = ht2+1
February 2, 2016
98 / 104
VaR in RiskMetrics
I
For the IGARCH(1,1) model under these assumptions, we have

rt [h ] N 0, ht2+1
Important assumptions:
I
I
I
I
I
normal errors: sum of normal r.v.s is normal.

no drift
volatility does not mean-revert (it has no mean)
suppose the tail probability is 5%

the daily VaR under RiskMetrics is :
VaR = Amount of Position 1.65t +1
the VaR at the h-day horizon under RiskMetrics is :
VaR (h ) = Amount of Position 1.65 ht +1

I
the square root of time rule...

February 2, 2016
99 / 104
Discussion of RiskMetrics
simple, makes risk transparent and tangible X
normality assumption is a weakness
square root of time rule is an artefact of the assumption that t = 0

I
in general:

VaR (h ) = Amount of Position 1.65 ht +1 h

VaR (h ) 6=
kVar (1)
February 2, 2016
100 / 104
Calculation of VaR from a general GARCH model

I
For many models, we do not know the CDF of the losses at horizon h.
We must use simulation.
Start at time t with known estimate t2 and mean t .

1. For i = 1, . . . , M, simulate future returns
(i )
(i )
(i )
rt +1 , rt +2 , . . . , rt +h
(i )
(i )
(i )
2,(i )
2,(i )
2,(i )
which requires simulating t +1 , t +2 , . . . , t +h and t +1 , t +2 , . . . , t +h

(i )
2. For each return sequence i, calculate the loss Lt,t +h at horizon h on the position.
n
oM
(i )
3. Sort the losses Lt,t +h
from smallest to largest. This is an approximation to
i =1
the CDF of the loss distribution.

4. As your estimator of VaR, take the nearest integer (1 p ) M in the sorted values
as the quantile.
The larger the value of M the more accurate the estimate.

February 2, 2016
101 / 104
VaR.
OneDay VaR for $1bn position
90
80
70
Millions of Dollars
60
50
40
30
20
10
0
1965
1971
1976
1982
1987
1993
1998
2004
2009
2015
EGARCH(1,1). Daily data. CRSP-VW.1971-2012. VaR for $1bn position in stocks.
February 2, 2016
102 / 104
Outline
1
Realized Variance
ARCH models
GARCH models
Value-at-Risk
References
February 2, 2016
103 / 104
Black, F. (1976).
Studies of stock market volatility changes.
In Proceedings of the American Statistical Association, Business and Economic Statistics Section, Volume 177, pp. 181.
Bollerslev, T. (1986).
Generalized autoregressive conditional heteroskedasticity.
Journal of Econometrics 31 (3), 307327.
Engle, R. F. (1982).
Autoregressive conditional heteroskedasticity with estimates of the variance of the United Kingdom inflation.
Econometrica 50 (4), 9871007.
Glosten, L. R., R. Jagannathan, and D. E. Runkle (1993).

On the relation between the expected value and the volatility of the nominal excess return on stocks.
The Journal of Finance 48 (5), 17791801.
Nelson, D. B. (1991).
Conditional heteroskedasticity in asset returns: a new approach.
Econometrica 59 (2), 347370.
February 2, 2016
104 / 104
Matlab
I
to specify Engles original ARCH (Q ) model, use the equivalent

GARCH (0, Q ) specification.
specify the ARCH (Q ) model as follows:

model = garch (0, Q )
math works link
estimate the ARCH (Q ) model as follows:
[EstMdl, EstParamCov , logL, info ] = estimate (Mdl, Y )

math works link
Back
February 2, 2016
104 / 104

Conditional Heteroskedasticity

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Conditional Heteroskedasticity

Transféré par

Droits d'auteur :

Formats disponibles

Time Series Analysis Stylized facts on Volatility Clustering Realized Variance ARCH models GARCH models I-GARCH, EGARCH

Lecture 5: Conditional Heteroskedasticity Models

Drew Creal (Chicago Booth)

Time Series Analysis

Stylized facts on Volatility Clustering

I-GARCH, EGARCH, GARCH-M, GJR models

Time Series Analysis

Stylized facts on Volatility Clustering

I-GARCH, EGARCH, GARCH-M, GJR models

so far, we have focused on modeling the conditional mean:

given estimates, 0 , 1 , . . . , q 1 , q , we can calculate the residuals

Drew Creal (Chicago Booth)

Why White Noise?

suppose the AR(1) is misspecified

then the forecast errors are correlated over time, i.e.:

forecast errors should be white noise

suppose we have a good model for the conditional mean t (e.g.

now we look at the squared residuals {2t } to test for conditional

for returns on most financial assets, there is a lot more

Drew Creal (Chicago Booth)

In most asset markets, volatility varies dramatically over time.

episodes of high volatility seem to be clustered.

we dont simply observe volatility

how do we measure vol?

Drew Creal (Chicago Booth)

Why do we care about Volatility?

why do we care about modeling vol?

Drew Creal (Chicago Booth)

ARCH model of Engle (1982)

the standard normal is a common choice for t

the (standardized) Students t is another option for t

large shocks are followed by more large shocks

Drew Creal (Chicago Booth)

Building a Volatility Model

1. estimate a model for the conditional mean (e.g. ARMA model)

Drew Creal (Chicago Booth)

Time Series Analysis

Stylized facts on Volatility Clustering

I-GARCH, EGARCH, GARCH-M, GJR models

Monthly Returns on Market Squared

log stock returns on market -squared. Monthly data. 1926-2012.

Drew Creal (Chicago Booth)

ACF for |log returns|

ACF for squared returns

PACF for squared returns

log stock returns on CRSP-VW. Monthly data. 1926-2012.

Drew Creal (Chicago Booth)

Dynamics of Variance-Monthly Returns

ACF for |log returns|

ACF for squared returns

PACF for squared returns

log stock returns on CRSP-VW. Monthly data. 1945-2012.

Drew Creal (Chicago Booth)

Dynamics of Variance-Monthly Returns

ACF for |log returns|

ACF for squared returns

PACF for squared returns

log stock returns on CRSP-VW. Monthly data. 1970-2012.

Drew Creal (Chicago Booth)

Daily Returns on Market Squared

log stock returns on market -squared. Daily data. 1988-2012.

Drew Creal (Chicago Booth)