Volatility

Many economic series, and most financial

series, display conditional volatility
The conditional variance changes over time
There are periods of high volatility
When large changes frequently occur
And periods of low volatility
When large changes are less frequent
 Weekly Stock Prices
Levels and Returns
1500

20
10
1000
sp

0
r
500

-10
0

-20
1950w1 1960w1 1970w1 1980w1 1990w1 2000w1 2010w1
t 1950w1 1960w1 1970w1 1980w1 1990w1 2000w1 2010w1
t
 Conditional Mean
The conditional mean of y is
E ( yt | t 1 )

The regression error is mean zero and

unforecastable
E (et | t 1 ) = 0
 Conditional Variance
The conditional variance of y is

(
var( yt | t 1 ) = E ( yt E ( yt | t 1 )) | t 1
2
)
(
= E e 2t | t 1 )
The squared regression error can be
forecastable
 Forecastable Conditional Variance
If the squared error is forecastable, then the
conditional variance is time-varying and
correlated.
The magnitude of changes is predictable
The sign is not predictable
 Stock returns are unpredictable
. reg r L(1/4).r,r

Linear regression Number of obs = 3120

F( 4, 3115) = 1.22
Prob > F = 0.3009
R-squared = 0.0033
Root MSE = 2.0793

Robust
r Coef. Std. Err. t P>|t| [95% Conf. Interval]

r
L1. -.01737 .032557 -0.53 0.594 -.0812053 .0464653
L2. .0466548 .0273551 1.71 0.088 -.0069811 .1002907
L3. -.0044898 .0282283 -0.16 0.874 -.0598378 .0508581
L4. -.0298138 .0264335 -1.13 0.259 -.0816427 .0220151

_cons .1337402 .0410395 3.26 0.001 .053273 .2142074

0.10
. testparm L(1/4).r

( 1) L.r = 0

0.05
Autocorrelations of r
( 2) L2.r = 0
( 3) L3.r = 0
( 4) L4.r = 0

F( 4, 3115) = 1.22 0.00

Prob > F = 0.3009
-0.05

0 10 20 30 40
Lag
Bartlett's formula for MA(q) 95% confidence bands
 Squared Returns are predictable
. gen y=(r-.1334364)^2
(1 missing value generated)

. reg y L(1/4).y,r

Linear regression Number of obs = 3120

F( 4, 3115) = 9.72
Prob > F = 0.0000
R-squared = 0.1092
Root MSE = 11.618

Robust
y Coef. Std. Err. t P>|t| [95% Conf. Interval]

y
L1. .2332184 .1248813 1.87 0.062 -.0116395 .4780763
L2. .0627729 .0308486 2.03 0.042 .0022873 .1232586
L3. .125441 .0364307 3.44 0.001 .0540103 .1968717
L4. .0517234 .0506949 1.02 0.308 -.0476755 .1511222

_cons 2.282243 .3480455 6.56 0.000 1.599821 2.964665

. testparm L(1/4).y

( 1) L.y = 0
( 2) L2.y = 0
( 3) L3.y = 0
( 4) L4.y = 0

F( 4, 3115) = 9.72
Prob > F = 0.0000
 Squared Returns

0.30
400

0.20
300

Autocorrelations of y
200

0.10
y
100

0.00 -0.10
0

1950w1 1960w1 1970w1 1980w1 1990w1 2000w1 2010w1

t 0 10 20 30 40
Lag
Bartlett's formula for MA(q) 95% confidence bands
 ARCH
Robert Engle (1982) proposed a
model for the conditional variance
AutoRegressive Conditional
Heteroskedasticity
ARCH now describes volatility
models
Nobel Prize 2003
 ARCH(1) Model
yt = + et
t2 = var(et | t 1 ) = + et21
>0
0
>0 means that the conditional variance is high when the
lagged squared error is high
Large errors (either sign) today mean high expected errors
(in magnitude) tomorrow.
Small magnitude errors forecast next period small
magnitude errors.
 Unconditional variance
A property of expectations is that expected
(average) conditional expectations are
unconditional expectations.
So the average conditional variance is the
average variance the variance of the
regression error.
2 = E ( t2 ) = + E (et21 ) = + 2
Solving for the variance: 2 =
1
 Rewriting, this implies
= (1 )
2

Substituting into ARCH(1) equation

t2 = (1 ) 2 + et21
or
= + (e
t
2 2 2
t 1
2
)
This shows that the conditional variance is a
combination of the unconditional variance, and
the deviation of the squared error from its
average value.
 ARCH(1) as AR(1) in squares
The model

( 2
t )
var(et | t 1 ) = E e | t 1 = + e2
t 1

implies the regression

e = + e
2
t
2
t 1 + ut
where u is white noise
Thus e-squared is an AR(1)
 Estimation
.arch r, arch(1)
ARCH family regression

Sample: 1950w2 - 2010w5 Number of obs = 3124

Distribution: Gaussian Wald chi2(.) = .
Log likelihood = -6525.268 Prob > chi2 = .

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .1996426 .0314495 6.35 0.000 .1380027 .2612825

ARCH
arch
L1. .3006982 .0216209 13.91 0.000 .2583219 .3430745

_cons 2.926873 .0686021 42.66 0.000 2.792416 3.061331

 Variance Forecast
Given the parameter estimates, the estimated
conditional variance for period t is

= + e = + ( yt 1 )
2 2 2
t t 1

The forecasted out-of-sample variance is

n +1 = + ( yn )
2
2
 Forecast Interval for the mean
You can use the estimated conditional
standard deviation to obtain forecast intervals
for the mean
y n +1|n Z / 2 n +1
These forecast intervals will vary in width
depending on the estimated conditional
variance.
Wider in periods of high volatility
More narrow in periods of low volatility
 ARCH(p) model
Allow p lags of squared errors
yt = + et
t2 = + 1et21 + 2 et2 2 + + p et2 p

Similar to AR(p) in squares

Estimation: ARCH(8)
.arch r, arch(1/8)
ARCH model with lags 1 through 8
 ARCH(8) Estimates
.arch r, arch(1/8)
ARCH family regression

Sample: 1950w2 - 2010w5 Number of obs = 3124

Distribution: Gaussian Wald chi2(.) = .
Log likelihood = -6368.552 Prob > chi2 = .

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .2027503 .0290226 6.99 0.000 .1458671 .2596335

ARCH
arch
L1. .1867283 .0163376 11.43 0.000 .1547071 .2187495
L2. .1099957 .0203355 5.41 0.000 .0701388 .1498526
L3. .1541191 .0225999 6.82 0.000 .1098241 .1984142
L4. .0912413 .0192753 4.73 0.000 .0534625 .1290202
L5. .0284588 .0171987 1.65 0.098 -.0052501 .0621677
L6. .0811242 .0192012 4.22 0.000 .0434906 .1187578
L7. .041083 .0147083 2.79 0.005 .0122553 .0699107
L8. .0706622 .0171741 4.11 0.000 .0370015 .1043229

_cons 1.144063 .1001062 11.43 0.000 .9478582 1.340267

 ARCH needs many lags
Notice that we included 8 lags, and all
appeared significant.
This is commonly observed in estimated ARCH
models
The conditional variance appears to be a function
of many lagged past squares
 GARCH Model
Tim Bollerslev (1986)
A student of Engle
Current faculty at Duke
proposed the GARCH model to simplify this
problem
t2 = + t21 + et21
>0
>0
0
 GARCH(1,1)
This makes the variance a function of all past
lags:
t = + t 1 + et 1
2 2 2

( )

= + e
j 2
t 1 j
j =0

It is also smoother than an ARCH model with a

small number of lags
 GARCH(p,q)
p lags of squared error
q lags of conditional variance

= + 1
t
2 2
t 1 + + q 2
t q + e2
1 t 1 ++ e2
p t p

GARCH(1,1):
.arch r, arch(1) garch(1)
GARCH(3,2):
.arch r, arch(1/3) garch(1/2)
ARCH family regression
GARCH(1,1)
Sample: 1950w2 - 2010w5 Number of obs = 3124
Distribution: Gaussian Wald chi2(.) = .
Log likelihood = -6359.118 Prob > chi2 = .

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .1935287 .0296365 6.53 0.000 .1354421 .2516152

ARCH
arch
L1. .1317621 .0094385 13.96 0.000 .113263 .1502613

garch
L1. .8444868 .0117076 72.13 0.000 .8215404 .8674333

_cons .1207574 .0221764 5.45 0.000 .0772924 .1642223

Common GARCH features

Lagged variance has large coefficient
Sum of two coefficients very close to (but less than) one
 GARCH(2,2) for Stock Returns
ARCH family regression

Sample: 1950w2 - 2010w5 Number of obs = 3124

Distribution: Gaussian Wald chi2(.) = .
Log likelihood = -6356.166 Prob > chi2 = .

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .1913999 .0295859 6.47 0.000 .1334126 .2493872

ARCH
arch
L1. .1658834 .0149416 11.10 0.000 .1365984 .1951684
L2. -.0277704 .042739 -0.65 0.516 -.1115373 .0559964

garch
L1. .5991681 .2941188 2.04 0.042 .0227059 1.17563
L2. .2373325 .2478461 0.96 0.338 -.248437 .7231021

_cons .1233949 .0428309 2.88 0.004 .0394478 .207342

 GARCH(1,1)
The GARCH(1,1) often fits well, and is a useful
benchmark.
Daily, weekly, or monthly asset returns, exchange
rates, or interest rates
 Extensions
There are many extensions of the basic
GARCH model, developed to handle a variety
of situations
Asymmetric Response
Garch-in-mean
Explanatory variables in variance
Non-normal errors
 Asymetric GARCH
Threshold GARCH
t2 = + t21 + et21 + et211(et 1 > 0)
The last tern is dummy variable for positive
lagged errors
This model specifies that the ARCH effect
depends on whether the error was positive or
negative
If the error is negative, the effect is
If the error is positive, the full effect is +
 TARCH estimation
.arch r, arch(1) tarch(1) garch(1)
Negative errors have coefficient of 0.19
Positive errors have coefficient of 0.05
Negative returns increase volatility much more than positive returns
ARCH family regression

Sample: 1950w2 - 2010w5 Number of obs = 3124

Distribution: Gaussian Wald chi2(.) = .
Log likelihood = -6332.433 Prob > chi2 = .

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .1474826 .0313132 4.71 0.000 .0861099 .2088552

ARCH
arch
L1. .1879679 .0154078 12.20 0.000 .1577692 .2181665

tarch
L1. -.1408097 .0160892 -8.75 0.000 -.1723439 -.1092754

garch
L1. .8437111 .0132294 63.78 0.000 .817782 .8696403

_cons .1540714 .0219836 7.01 0.000 .1109843 .1971585

 Leverage Effect
This model describes what is called the
leverage effect
A negative shock to equity increases the ratio
debt/equity of investors
This increases the leverage of their portfolios
This increases risk, and the conditional variance
Negative shocks have stronger effect on variance
than positive shocks
 GARCH-in-mean
If investors are risk averse, risky assets will
earn higher returns (a risk premium) in market
equilibrium
If assets have varying volatility (risk), their
expected return will vary with this volatility
Expected return should be positively correlated
with volatility
 GARCH-M model

y = 1 + 1 t21 + et
t2 = + t21 + et21
.arch arch(1) garch(1) archm
 GARCH-M for Stock Returns
Marginally positive effect
ARCH family regression

Sample: 1950w2 - 2010w5 Number of obs = 3124

Distribution: Gaussian Wald chi2(1) = 2.96
Log likelihood = -6357.259 Prob > chi2 = 0.0853

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .1211625 .0512489 2.36 0.018 .0207165 .2216085

ARCHM
sigma2 .024739 .0143783 1.72 0.085 -.003442 .0529199

ARCH
arch
L1. .1315334 .0096454 13.64 0.000 .1126287 .1504381

garch
L1. .8450762 .0118319 71.42 0.000 .8218862 .8682662

_cons .1193442 .022376 5.33 0.000 .075488 .1632004

 TARCH and GARCH-M
.arch arch(1) tarch(1) garch(1) archm
archm term appears insignificant
ARCH family regression

Sample: 1950w2 - 2010w5 Number of obs = 3124

Distribution: Gaussian Wald chi2(1) = 0.17
Log likelihood = -6332.324 Prob > chi2 = 0.6776

OPG
r Coef. Std. Err. z P>|z| [95% Conf. Interval]

r
_cons .1309694 .0512904 2.55 0.011 .030442 .2314967

ARCHM
sigma2 .0059058 .0142042 0.42 0.678 -.021934 .0337456

ARCH
arch
L1. .1872044 .0155283 12.06 0.000 .1567694 .2176393

tarch
L1. -.1391533 .0161185 -8.63 0.000 -.1707449 -.1075617

garch
L1. .8425617 .0137719 61.18 0.000 .8155693 .8695542

_cons .1565728 .0228305 6.86 0.000 .1118259 .2013197

 Estimated standard deviation
Estimated TARCH model
.predict v, variance
.gen s=sqrt(v)

10

Unconditional
8

variance is 2.1
6
s
4
2
0

1950w1 1960w1 1970w1 1980w1 1990w1 2000w1 2010w1

t
 S&P, returns, and standard deviation
2006-2010
1600

10
1400
1200

8
sp
1000
800

6
s
600

2006w1 2007w1 2008w1 2009w1 2010w1

t

4
10

2
0

2006w1 2007w1 2008w1 2009w1 2010w1

r

t
-10
-20

2006w1 2007w1 2008w1 2009w1 2010w1

t