Vol Modeling (ARCH, GARCH, LR Etc.)

Volatility Modeling
Lecture 9: Volatility Modeling
MIT 18.S096
Dr. Kempthorne
Fall 2013
MIT 18.S096 Volatility Modeling 1

Defining Volatility
Historical Volatility: Measurement and Prediction
Geometric Brownian Motion
Volatility Modeling
Poisson Jump Diffusions
ARCH Models
GARCH Models
Outline
1 Volatility Modeling
Defining Volatility
ARCH Models
GARCH Models

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Defining Volatility
Basic Definition
Annualized standard deviation of the change in price or value
of a financial security.
Estimation/Prediction Approaches
Historical/sample volatility measures.
Geometric Brownian Motion Model
Poisson Jump Diffusion Model
ARCH/GARCH Models
Stochastic Volatility (SV) Models
Implied volatility from options/derivatives

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Outline
Defining Volatility
ARCH Models
GARCH Models

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Historical Volatility
Computing volatility from historical series of actual prices
Prices of an asset at (T + 1) time points
{Pt , t = 0, 1, 2, . . . , T }
Returns of the asset for T time periods
Rt = log (Pt /Pt1 ), t = 1, 2, . . . , T
{Rt } assumedpcovariance stationary
p with
= var (Rt ) = E [(Rt E [Rt ])2 ]
with sample estimate:
q PT
= T 11 t=1 (Rt R)2 , with R = T1 T
P
1 Rt .
Annualized values
252 (daily prices for 252 business days/year)
vol =
c 52 (weekly prices)

12 (monthly prices)
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Prediction Methods Based on Historical Volatility

Definition For time period t, define the sample volatility
t = sample standard deviation of period t returns
If t indexes months with daily data, then t is the sample standard deviation of
daily returns in month t.
If t indexes days with daily data, then t2 = Rt2 .
With high-frequency data, daily t is derived from cumulating squared intra-day
returns.
2 = 1t t1 j2
P
Historical Average: t+1
(uses all available data)
2 = m1 m1 2
P
Simple Moving Average: t+1 0 tj
(uses last m single-period sample estimates)
Exponential Moving Average: t+1 2 = (1 )t2 + t2 0 1
(uses all available data)
Exponential WeightedPMoving Average:
1 j 2 Pm1 j
t+1 = m
2
j=0 ( t j )/[ j=0 ] (uses last m)
single-period sample estimates).
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Predictions Based on Historical Volatility

Simple Regression:
2
t+1 = 1,t t2 + 1,t t1
2 2
+ + p,t tp+1 + ut
Regression can be fit using all data or last m (rolling-windows).
Note: similar but different from auto-regression model of t2
Trade-Offs
Use more data to increase precision of estimators
Use data closer to time t for estimation of t .
Evaluate out-of-sample performance
Distinguish assets and asset-classes
Consider different sampling frequencies and different forecast
horizons
Apply performance measures (MSE, MAE, MAPE, etc.)
Benchmark Methodology: RiskMetrics
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Outline
Defining Volatility
ARCH Models
GARCH Models

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Geometric Brownian Motion (GBM)

For {S(t)} the price of a security/portfolio at time t:
dS(t) = S(t)dt + S(t)dW (t),
where
is the volatility of the securitys price
is mean return (per unit time).
dS(t) infinitesimal increment in price
dW (t)infinitesimal increment of a standard Brownian
Motion/Wiener Process
Increments [W (t 0 ) W (t)] are Gaussian with mean zero and
variance (t 0 t).
Increments on disjoint time intervals are independent.
For t1 < t2 < t3 < t4 ,
[W (t2 ) W (t1 )] and [W (t4 ) W (t3 )] are independent
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Geometric Brownian Motion (GBM)

Sample Data from Process:
Prices: {S(t), t = t0 , t1 , . . . , tn }
Returns: {Rj = log [S(tj )/S(tj 1 )], j = 1, 2, . . . , n}
indep. r.v.s: Rj N( j , 2 j ), where
j = (tj tj1 ) and = [ 2 /2]
({log [S(t)]} is Brownian Motion with drift and volatility 2 .)
Maximum-Likelihood Parameter Estimation

If j 1, then
= n1 n1 Rt
P
= R P
2 = n1 n1 (Rt R)2
If j varies ... Exercise.
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Garman-Klass Estimator:
Sample information more than period-close prices, also have
period-high, period-low, and period-open prices.
Assume = 0, j 1 (e.g., daily) and let f (0, 1) denote
the fraction of the day prior to the market open.
Cj = log [S(tj )]
Oj = log [S(tj1 + f )]
Hj = max log [S(t)]

tj 1 +f ttj
Lj = min log [S(t)]

tj1 +f ttj

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Garman-Klass Estimator
Using data from the first period:
02 = (C1 C0 )2 : Close-to-Close squared return
E [02 ] = 2 , and var [02 ] = 2( 2 )2 = 2 4 .
(O1 C0 )2
12 = f : Close-to-Open squared return
E [1 ] = 2 , and var [12 ] = 2( 2 )2 = 2 4 .
2
O1 )2
22 = (C11f : Open-to-Close squared return
E [2 ] = 2 , and var [22 ] = 2( 2 )2 = 2 4 .
2
Note: 1 and 22 are independent!

2
2 = 12 12 + 21 22
E [2 ] = 2 , and var [2 ] = 4 .
var (02 )
= eff (2 ) = var (2 )
= 2.
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Parkinson (1976): With f = 0, defines

2
1 L1 )
32 = (H4(log 2
2) and shows eff (3 ) 5.2.
Garman and Klass (1980) show that for any 0 < f < 1:
42 = a 12 + (1 a)32
has minimum variance when a 0.17, independent of f and
Eff (42 ) 6.2.
Best Analytic Scale-Invariant Estimator
2 = 0.511(u d )2 0.019{c (u + d ) 2u d } 0.383c 2 ,
1 1 1 1 1 1 1 1
where the normalized high/low/close are:
uj = Hj Oj
dj = Lj Oj
cj = Cj Oj
and 2 ) 7.4
Eff (

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
If 0 < f < 1 then the opening price O1 may differ from C0

and the composite estimator is
2 2
2 = a (O1 C0 ) + (1 a)
GK f (1f )
which has minimum variance when a = 0.12 and
2 ) 8.4.
Eff (GK

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Outline
Defining Volatility
ARCH Models
GARCH Models

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models

For {S(t)} the stochastic proces for the price of the
security/portfolio at time t,
dS(t)
S(t) = dt + dW (t) + Z (t)d(t),
where
dS(t) = infinitesimal increment in price.
= mean return (per unit time)
= diffusion volatility of the securitys price process
dW (t)= increment of standard Wiener Process
d(t) = increment of a Poisson Process with rate ,
modeling the jump process.
() Z (t), the magnitude a return jump/shock
Z (t) i.i.d N(0, 1) r.v.s and
= scale( units) of jump magnitudes.
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Maximum-Likelihood Estimation of the PJD Model

Model is a Poisson mixture of Gaussian Distributions.
Moment-generating function derived as that of random sum
of independent random variables.
Likelihood function product of infinite sums
EM Algorithm* expressible in closed form
Jumps treated as latent variables which simplify computations
Algorithm provides a posteriori estimates of number of jumps
per time period.
* See Pickard, Kempthorne, Zakaria (1987).

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Outline
Defining Volatility
ARCH Models
GARCH Models

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
ARCH Models
ARCH models are specified relative to the discrete-time process for

the price of the security/portfolio: {St , t = 1, 2, . . .}
Engle (1982) models the discrete returns of the process
yt = log (St /St1 ) as
y t = t + t ,
where t is the mean return, conditional on Ft1 , the information
available through time (t 1), and
t = Z t t ,
where Zt i.i.d. with E [Zt ] = 0, and var [Zt ] = 1,
t2 = 0 + 1 2t1 + 2 2t2 + + p 2tp
Parameter Constraints: j 0, j = 0, 1, . . . , p
t2 = var (Rt | Ft1 ), Conditional Heteroscedasticity of returns .

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
ARCH Models
The ARCH model:
t2 = 0 + 1 2t1 + 2 2t2 + + p 2tp
implies an AR model in 2t . Add (2t t2 ) = ut to both sides:
2t = 0 + 1 2t1 + 2 2t2 + + p 2tp + ut
where ut : E [ut | Ft ] = 0, and var [ut | Ft ] = var (2t ) = 2t4 .
Lagrange Multiplier Test
H0 : 1 = 2 = = p = 0
Fit linear regression on squared residuals t = yt t .
(i.e., Fit an AR(p) model to [2t ], t = 1, 2, . . . , n)
LM test statistic = nR 2 , where R 2 is the R-squared of the
fitted AR(p) model.
Under H0 the r.v. nR 2 is approx. 2 (df = p)
Note: the linear regression estimates of parameters are not MLEs under Gaussian
assumptions; they correspond to quasi-maximum likelihood estimates (QMLE).
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Maximum Likelihood Estimation

ARCH Model:
yt = c + t
t = zt t
t2 = 0 + 1 2t1 + + p 2tp
t = 0, 1, . . . , T
Likelihood:
L(c, ) = pQ(yn 1 , . . . , yn | c, 0 , 1 , . . . , p )
= p(yt | Ft1 , c, )
Qt=1
n 2t
= t=1 [
1 2 exp( 1 2 2 )]
2t t
where t = yt c and
t2 = 0 + 1 2t1 + p 2tp .
Constraints:
i 0, i = 1, 2, . . . , p
(1 + + p ) < 1.
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Outline
Defining Volatility
ARCH Models
GARCH Models

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
GARCH Models
Bollerslev (1986) extended ARCH models to:

GARCH(p,q) Model
Pp Pq
t2 = 0 + i=1 i 2ti + j=1 2
j tj
Constraints: i 0, i, and j 0, j
GARCH(1,1) Model
t2 = 0 + 1 2t1 + 1 t1
2
Parsimonious
Fits many financial time seriies

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
GARCH Models
The GARCH(1,1) model:
t2 = 0 + 1 2t1 + 1 t21
implies an ARMA model in 2t . Eliminate t20 using (2t 0 t20 ) = ut 0
2t ut = 0 + 1 + 2t1 + 1 (2t1 ut1 )
2t = 0 + (1 + 1 )2t1 + ut 1 ut1
where ut : E [ut | Ft ] = 0, and var [ut | Ft ] = var (2t ) = 2t4 .
= GARCH(1, 1) implies an ARMA(1, 1) with
ut = (2t t2 ) WN(0, 2 4 )
Stationarity of GARCH model deduced from ARMA model
A(L)2t = B(L)ut
2t = [A(L)]1 B(L)ut .
Covariance stationary:roots of A(z) outside {|z| 1},i.e.,
|1 + 1 | < 1
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
GARCH Models
Unconditional Volatility / Long-Run Variance

GARCH(1,1): Assuming stationarity, 0 < (1 + 1 ) < 1
2 = 0 + (1 + 1 )2
0
= 2 = (11 1 )
GARCH(p,q) implies ARMA( max(p,q), q) model
Stationary if 0 < ( p1 i + q1 j ) < 1
P P
Long-Run Variance:
2 = 0 + ( p1 j + q1 1 )2
P P
0
= 2 = Pmax(p,q)
[1 1 (i +j )]

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
GARCH Model Estimation

Maximum Likelihood Estimation
GARCH Model:
yt = c + t
t = zt t P
t2 = 0 + 1p i 2ti + 1q j t2j
P
t = 0, 1, . . . , T
Likelihood:
L(c, , ) = p (y1 , . . . , yT | c, 0 , 1 , . . . , p , 1 , . . . , q )
QT
= p(yt | Ft 1 , c, , )
Qt=1
T 2t
= t=1 [
1 2 exp( 1 2 2 )]
2t t
where t = yt cPand
t2 = 0 + p1 i 2ti + qj=1 j t2j .
P
Constraints: i 0, i, j 0, j, and 0 < ( p1 u + 11 j ) < 1.
P P

Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
GARCH Model Estimation/Evaluation/Model-Selection

Maximum-Likelihood Estimates: c, ,
= t and t2 (t = T , T 1, . . .)
Standardized Residuals
t /t : should be uncorrelated
Squared Standardized Residuals
(t /t )2 : should be uncorrelated
Testing Normality of Residuals
Normal QQ Plots
Jaque-Bera test
Shapiro-Wilk test
MLE Percentiles Goodness-of-Fit Test
Kolmogorov-Smirnov Goodness-of-Fit Test
Model Selection: Apply model-selection critera
Akaike Information Criterion (AIC)
Bayes Information Criterion (BIC)
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Stylized Features of Returns/Volatility

Volatility Clustering
Large 2t follow large 2t1
Small 2t follow small 2t1
GARCH models can prescribe
Large t2 follow large t1
2
2 2
Small t follow small t1
Heavy Tails / Fat Tails
Returns distribution has heavier tails (higher Kurtosis) than
Gaussian
GARCH(p,q) models are stochastic mixture of Gaussian
distributions with higher kurtosis.
Engle, Bollerslev, and Nelson (1994)
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Stylized Features of Returns/Volatility
Volatility Mean Reversion

GARCH(1,1) Model
Long-run average volatility: 2 = 1101
Mean-Reversion to Long-Run Average
2t = 0 + (1 + 1 )2t1 + ut 1 ut1
Substituting: 0 = (1 1 1 )2
(2t 2 ) = (1 + 1 )(2t 1 2 ) + ut 1 ut1
0 < (1 + 1 ) < 1 = Mean Reversion!
Extended Ornstein-Uhlenbeck(OU) Process for

2t with MA(1) errors.
Defining Volatility
Volatility Modeling
ARCH Models
GARCH Models
Extensions of GARCH Models
EGARCH Nelson (1992)

TGARCH Glosten, Jagannathan, Runkler (1993)
PGARCH Ding, Engle, Granger
GARCH-In-Mean
Non-Gaussian Distributions
tDistributions
Generalized Error Distributions (GED)

MIT OpenCourseWare
http://ocw.mit.edu
18.S096 Topics in Mathematics with Applications in Finance

Fall 2013
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

Vol Modeling (ARCH, GARCH, LR Etc.)

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Vol Modeling (ARCH, GARCH, LR Etc.)

Uploaded by

Copyright:

Available Formats

Volatility Modeling

Lecture 9: Volatility Modeling

MIT 18.S096 Volatility Modeling 1

MIT 18.S096 Volatility Modeling 2

MIT 18.S096 Volatility Modeling 3

MIT 18.S096 Volatility Modeling 4

Prediction Methods Based on Historical Volatility

Predictions Based on Historical Volatility

MIT 18.S096 Volatility Modeling 8

Geometric Brownian Motion (GBM)

Geometric Brownian Motion (GBM)

Maximum-Likelihood Parameter Estimation

Geometric Brownian Motion

Hj = max log [S(t)]

Lj = min log [S(t)]

MIT 18.S096 Volatility Modeling 11

Note: 1 and 22 are independent!

Parkinson (1976): With f = 0, defines

MIT 18.S096 Volatility Modeling 13

If 0 < f < 1 then the opening price O1 may differ from C0

MIT 18.S096 Volatility Modeling 14

MIT 18.S096 Volatility Modeling 15

Poisson Jump Diffusions

Poisson Jump Diffusions

Maximum-Likelihood Estimation of the PJD Model

MIT 18.S096 Volatility Modeling 17

MIT 18.S096 Volatility Modeling 18

ARCH models are specified relative to the discrete-time process for

MIT 18.S096 Volatility Modeling 19

Maximum Likelihood Estimation

MIT 18.S096 Volatility Modeling 22

Bollerslev (1986) extended ARCH models to:

MIT 18.S096 Volatility Modeling 23

Unconditional Volatility / Long-Run Variance

MIT 18.S096 Volatility Modeling 25

GARCH Model Estimation

MIT 18.S096 Volatility Modeling 26

GARCH Model Estimation/Evaluation/Model-Selection

Stylized Features of Returns/Volatility

Stylized Features of Returns/Volatility

Volatility Mean Reversion

0 < (1 + 1 ) < 1 = Mean Reversion!

Extended Ornstein-Uhlenbeck(OU) Process for

Extensions of GARCH Models

EGARCH Nelson (1992)

MIT 18.S096 Volatility Modeling 30

18.S096 Topics in Mathematics with Applications in Finance

You might also like