Modelling Volatility and Correlation

Chapter 9
Modelling volatility and correlation
‘Introductory Econometrics for Finance’ © Chris Brooks 2013 1

An Excursion into Non-linearity Land
• Motivation: the linear structural (and time series) models cannot

explain a number of important features common to much financial data
- leptokurtosis
- volatility clustering or volatility pooling
- leverage effects
• Our “traditional” structural model could be something like:

yt = 1 + 2x2t + ... + kxkt + ut, or more compactly y = X + u
• We also assumed that ut  N(0,2).

A Sample Financial Asset Returns Time Series
Daily S&P 500 Returns for August 2003 – August 2013

Non-linear Models: A Definition
• Campbell, Lo and MacKinlay (1997) define a non-linear data

generating process as one that can be written
yt = f(ut, ut-1, ut-2, …)
where ut is an iid error term and f is a non-linear function.
• They also give a slightly more specific definition as

yt = g(ut-1, ut-2, …)+ ut2(ut-1, ut-2, …)
where g is a function of past error terms only and 2 is a variance
term.
• Models with nonlinear g(•) are “non-linear in mean”, while those with
nonlinear 2(•) are “non-linear in variance”.

Types of non-linear models
• The linear paradigm is a useful one.

• Many apparently non-linear relationships can be made linear by a
suitable transformation.
• On the other hand, it is likely that many relationships in finance are
intrinsically non-linear.
• There are many types of non-linear models, e.g.

- ARCH / GARCH
- switching models
- bilinear models

Testing for Non-linearity – The RESET Test
• The “traditional” tools of time series analysis (acf’s, spectral analysis)

may find no evidence that we could use a linear model, but the data
may still not be independent.
• Portmanteau tests for non-linear dependence have been developed.
• The simplest is Ramsey’s RESET test, which took the form:
ut  0  1 yt2  2 yt3 ...  p 1 ytp  vt
• Here the dependent variable is the residual series and the independent
variables are the squares, cubes, …, of the fitted values.

Testing for Non-linearity – The BDS Test
• Many other non-linearity tests are available - e.g., the BDS and
bispectrum test
• BDS is a pure hypothesis test. That is, it has as its null hypothesis that
the data are pure noise (completely random)
• It has been argued to have power to detect a variety of departures from
randomness – linear or non-linear stochastic processes, deterministic
chaos, etc)
• The BDS test follows a standard normal distribution under the null
• The test can also be used as a model diagnostic on the residuals to ‘see
what is left’
• If the proposed model is adequate, the standardised residuals should be
white noise.
Chaos Theory
• Chaos theory is a notion taken from the physical sciences

• It suggests that there could be a deterministic, non-linear set of
equations underlying the behaviour of financial series or markets
• Such behaviour will appear completely random to the standard
statistical tests
• A positive sighting of chaos implies that while, by definition, long-
term forecasting would be futile, short-term forecastability and
controllability are possible, at least in theory, since there is some
deterministic structure underlying the data
• Varying definitions of what actually constitutes chaos can be found in
the literature.

Detecting Chaos
• A system is chaotic if it exhibits sensitive dependence on initial

conditions (SDIC)
• So an infinitesimal change is made to the initial conditions (the initial
state of the system), then the corresponding change iterated through the
system for some arbitrary length of time will grow exponentially
• The largest Lyapunov exponent is a test for chaos
• It measures the rate at which information is lost from a system
• A positive largest Lyapunov exponent implies sensitive dependence,
and therefore that evidence of chaos has been obtained
• Almost without exception, applications of chaos theory to financial
markets have been unsuccessful
• This is probably because financial and economic data are usually far
noisier and ‘more random’ than data from other disciplines
Neural Networks
• Artificial neural networks (ANNs) are a class of models whose

structure is broadly motivated by the way that the brain performs
computation
• ANNs have been widely employed in finance for tackling time series
and classification problems
• Applications have included forecasting financial asset returns,
volatility, bankruptcy and takeover prediction
• Neural networks have virtually no theoretical motivation in finance
(they are often termed a ‘black box’)
• They can fit any functional relationship in the data to an arbitrary
degree of accuracy.

Feedforward Neural Networks
• The most common class of ANN models in finance are known as

feedforward network models
• These have a set of inputs (akin to regressors) linked to one or more
outputs (akin to the regressand) via one or more ‘hidden’ or
intermediate layers
• The size and number of hidden layers can be modified to give a closer
or less close fit to the data sample
• A feedforward network with no hidden layers is simply a standard
linear regression model
• Neural network models work best where financial theory has virtually
nothing to say about the likely functional form for the relationship
between a set of variables.

Neural Networks – Some Disadvantages
• Neural networks are not very popular in finance and suffer from
several problems:
– The coefficient estimates from neural networks do not have any real
theoretical interpretation
– Virtually no diagnostic or specification tests are available for estimated
models
– They can provide excellent fits in-sample to a given set of ‘training’ data,
but typically provide poor out-of-sample forecast accuracy
– This usually arises from the tendency of neural networks to fit closely to
sample-specific data features and ‘noise’, and so they cannot ‘generalise’
– The non-linear estimation of neural network models can be cumbersome
and computationally time-intensive, particularly, for example, if the model
must be estimated repeatedly when rolling through a sample.

Models for Volatility
• Modelling and forecasting stock market volatility has been the subject
of vast empirical and theoretical investigation
• There are a number of motivations for this line of inquiry:
– Volatility is one of the most important concepts in finance
– Volatility, as measured by the standard deviation or variance of returns, is
often used as a crude measure of the total risk of financial assets
– Many value-at-risk models for measuring market risk require the
estimation or forecast of a volatility parameter
– The volatility of stock market prices also enters directly into the Black–
Scholes formula for deriving the prices of traded options
• We will now examine several volatility models.

Historical Volatility
• The simplest model for volatility is the historical estimate

• Historical volatility simply involves calculating the variance (or
standard deviation) of returns in the usual way over some historical
period
• This then becomes the volatility forecast for all future periods
• Evidence suggests that the use of volatility predicted from more
sophisticated time series models will lead to more accurate forecasts
and option valuations
• Historical volatility is still useful as a benchmark for comparing the
forecasting ability of more complex time models

Heteroscedasticity Revisited
• An example of a structural model is

yt = 1 + 2x2t + 3x3t + 4x4t + u t
with ut  N(0, u2 ).
• The assumption that the variance of the errors is constant is known as

homoscedasticity, i.e. Var (ut) =  u2 .
• What if the variance of the errors is not constant?

- heteroscedasticity
- would imply that standard error estimates could be wrong.
• Is the variance of the errors likely to be constant over time? Not for
financial data.
Autoregressive Conditionally Heteroscedastic
(ARCH) Models
• So use a model which does not assume that the variance is constant.
• Recall the definition of the variance of ut:
t2= Var(ut ut-1, ut-2,...) = E[(ut-E(ut))  ut-1, ut-2,...]
2
We usually assume that E(ut) = 0

so  2= Var(ut  ut-1, ut-2,...) = E[ut2 ut-1, ut-2,...].
t
What could the current value of the variance of the errors plausibly
depend upon?
– Previous squared error terms.
• This leads to the autoregressive conditionally heteroscedastic model
for the variance of the errors:
t2= 0 + 1ut21
• This is known as an ARCH(1) model
• The ARCH model due to Engle (1982) has proved very useful in
finance.
Autoregressive Conditionally Heteroscedastic
(ARCH) Models (cont’d)
• The full model would be
yt = 1 + 2x2t + ... + kxkt + ut , ut  N(0, t2)
where t2 = 0 + 1 ut21
• We can easily extend this to the general case where the error variance
depends on q lags of squared errors:
t2 = 0 + 1 ut 1 +2 ut  2 +...+qut  q
2 2 2
• This is an ARCH(q) model.
• Instead of calling the variance t2, in the literature it is usually called ht,
so the model is
yt = 1 + 2x2t + ... + kxkt + ut , ut  N(0,ht)
where ht = 0 + 1 ut21 +2 ut2 2 +...+q ut2 q

Another Way of Writing ARCH Models
• For illustration, consider an ARCH(1). Instead of the above, we can

write
yt = 1 + 2x2t + ... + kxkt + ut , ut = vtt
t  0  1ut21 , vt  N(0,1)
• The two are different ways of expressing exactly the same model. The
first form is easier to understand while the second form is required for
simulating from an ARCH model, for example.

Testing for “ARCH Effects”
1. First, run any postulated linear regression of the form given in the equation
above, e.g. yt = 1 + 2x2t + ... + kxkt + ut
saving the residuals, ût .
2. Then square the residuals, and regress them on q own lags to test for ARCH
of order q, i.e. run the regression
uˆt2   0   1uˆt21   2uˆt2 2  ...   quˆt2 q  vt
where vt is iid.
Obtain R2 from this regression
3. The test statistic is defined as TR2 (the number of observations multiplied

by the coefficient of multiple correlation) from the last regression, and is
distributed as a 2(q).

Testing for “ARCH Effects” (cont’d)
4. The null and alternative hypotheses are

H0 : 1 = 0 and 2 = 0 and 3 = 0 and ... and q = 0
H1 : 1  0 or 2  0 or 3  0 or ... or q  0.
If the value of the test statistic is greater than the critical value from the
2 distribution, then reject the null hypothesis.
• Note that the ARCH test is also sometimes applied directly to returns
instead of the residuals from Stage 1 above.

Problems with ARCH(q) Models
• How do we decide on q?
• The required value of q might be very large
• Non-negativity constraints might be violated.
– When we estimate an ARCH model, we require i >0  i=1,2,...,q
(since variance cannot be negative)
• A natural extension of an ARCH(q) model which gets around some of

these problems is a GARCH model.

Generalised ARCH (GARCH) Models
• Due to Bollerslev (1986). Allow the conditional variance to be dependent

upon previous own lags
• The variance equation is now
t2 = 0 + 1 ut21 +t-12 (1)
• This is a GARCH(1,1) model, which is like an ARMA(1,1) model for the
variance equation.
• We could also write
t-12 = 0 + 1ut22 +t-22
t-22 = 0 + 1 ut23 +t-32
• Substituting into (1) for t-12 :
t2 = 0 + 1 ut21 +(0 + 1ut22 + t-22)
= 0 + 1 ut21 +0 + 1ut22 + t-22
Generalised ARCH (GARCH) Models (cont’d)
• Now substituting into (2) for t-22

t2 = 0 + 1 ut21 +0 + 1ut22 +2(0 + 1ut23 +t-32)
t2 = 0 + 1 ut21 +0 + 1ut22 +02 + 12ut23 +3t-32
t2 = 0 (1++2) + 1 ut21 (1+L+2L2 ) + 3t-32
• An infinite number of successive substitutions would yield
t2 = 0 (1++2+...) + 1 ut21 (1+L+2L2+...) + 02
• So the GARCH(1,1) model can be written as an infinite order ARCH model.
• We can again extend the GARCH(1,1) model to a GARCH(p,q):

t2 = 0+1 ut21 +2ut2 2 +...+qut2q +1t-12+2t-22+...+pt-p2
q p
t2 =  0   i u    j t  j
2 2
t i
i 1 j 1

Generalised ARCH (GARCH) Models (cont’d)
• But in general a GARCH(1,1) model will be sufficient to capture the

volatility clustering in the data.
• Why is GARCH Better than ARCH?

- more parsimonious - avoids overfitting
- less likely to breech non-negativity constraints

The Unconditional Variance under the GARCH
Specification
• The unconditional variance of ut is given by

0
Var(ut) =
1  (1   )
when 1   < 1
• 1    1 is termed “non-stationarity” in variance
• 1   = 1 is termed intergrated GARCH
• For non-stationarity in variance, the conditional variance forecasts will

not converge on their unconditional value as the horizon increases.

Estimation of ARCH / GARCH Models
• Since the model is no longer of the usual linear form, we cannot use
OLS.
• We use another technique known as maximum likelihood.
• The method works by finding the most likely values of the parameters
given the actual data.
• More specifically, we form a log-likelihood function and maximise it.

Estimation of ARCH / GARCH Models (cont’d)
• The steps involved in actually estimating an ARCH or GARCH model

are as follows
1. Specify the appropriate equations for the mean and the variance - e.g. an
AR(1)- GARCH(1,1) model:
yt =  + yt-1 + ut , ut  N(0,t2)
t2 = 0 + 1 ut21 +t-12
2. Specify the log-likelihood function to maximise:
T 1 T 1 T
L   log( 2 )   log(  t )   ( yt    yt 1 ) 2 /  t
2 2
2 2 t 1 2 t 1
3. The computer will maximise the function and give parameter values and
their standard errors
Parameter Estimation using Maximum Likelihood
• Consider the bivariate regression case with homoscedastic errors for

simplicity: y t   1   2 xt  u t
• Assuming that ut  N(0,2), then yt  N( 1   2 xt , 2) so that the

probability density function for a normally distributed random variable
with this mean and variance is given by
1  1 ( y t   1   2 xt ) 2  (1)
f ( y t  1   2 xt ,  ) 
2
exp  
 2  2 2 
• Successive values of yt would trace out the familiar bell-shaped curve.
• Assuming that ut are iid, then yt will also be iid.

(cont’d)
• Then the joint pdf for all the y’s can be expressed as a product of the
individual density functions
f ( y1 , y 2 ,..., yT  1   2 X t ,  2 )  f ( y1  1   2 X 1 ,  2 ) f ( y 2  1   2 X 2 ,  2 )...
f ( yT  1   2 X 4 ,  2 )
(2)
T
  f ( yt 1   2 X t ,  2 )
t 1
• Substituting into equation (2) for every yt from equation (1),

1  1 T ( y t   1   2 xt ) 2 
f ( y1 , y 2 ,..., yT 1   2 xt ,  )  T
2
exp    (3)
 ( 2 ) T
 2 t 1  2

(cont’d)
• The typical situation we have is that the xt and yt are given and we want to
estimate 1, 2, 2. If this is the case, then f() is known as the likelihood
function, denoted LF(1, 2, 2), so we write
1  1 T ( y t   1   2 xt ) 2 
LF ( 1 ,  2 ,  )  T
2
exp    (4)
 ( 2 ) T
 2 t 1  2

• Maximum likelihood estimation involves choosing parameter values (1,

2,2) that maximise this function.
• We want to differentiate (4) w.r.t. 1, 2,2, but (4) is a product containing
T terms.

(cont’d)
• Since max f ( x )  maxlog( f ( x )) , we can take logs of (4).
x x
• Then, using the various laws for transforming functions containing

logarithms, we obtain the log-likelihood function, LLF:
T 1 T ( y t   1   2 xt ) 2
LLF  T log   log( 2 )  
2 2 t 1 2
• which is equivalent to
T T 1 T ( y t   1   2 xt ) 2
LLF   log   log( 2 )  
2
(5)
2 2 2 t 1 2
• Differentiating (5) w.r.t. 1, 2,2, we obtain

LLF 1 ( y   1   2 xt ).2.  1
  t
 1 2 2 (6)
(cont’d)
LLF 1 ( yt  1   2 xt ).2.  xt (7)
 
 2 2 2
LLF T 1 1 ( y t   1   2 xt ) 2
   (8)
 2 22 2 4
• Setting (6)-(8) to zero to minimise the functions, and putting hats above
the parameters to denote the maximum likelihood estimators,
• From (6),  ( y  ˆ  ˆ x )  0
t 1 2 t
 y   ˆ   ˆ x  0
t 1 2 t
 y  Tˆ  ˆ  x  0
t 1 2 t
1 ˆ  ˆ 1
T
 t 1 2 T  xt  0
y  
(9)
ˆ1  y  ˆ 2 x
(cont’d)
• From (7),  ( y  ˆ  ˆ x ) x  0
t 1 2 t t
 y x   ˆ x   ˆ x  0
t t 1 t
2
2 t
 y x  ˆ  x  ˆ  x  0
t t 1 t 2
2
t
ˆ  x   y x  ( y  ˆ x ) x
2
2
t t t 2 t
ˆ  x   y x  Txy  ˆ Tx
2
2
t t t 2
2
ˆ 2 ( xt2 Tx 2 )   yt xt  Txy
ˆ 2 
 y x  Txy
t t
(10)
( x Tx ) 2
t
2
T 1
• From (8), ˆ 2


ˆ 4
(y t  ˆ1  ˆ 2 xt ) 2
(cont’d)
• Rearranging, ˆ 2  1  ( y t  ˆ1  ˆ 2 xt ) 2
T
1
 2   ut2 (11)
T
• How do these formulae compare with the OLS estimators?

(9) & (10) are identical to OLS
(11) is different. The OLS estimator was
1
 2 
Tk
 ut2
• Therefore the ML estimator of the variance of the disturbances is biased,

although it is consistent.
• But how does this help us in estimating heteroscedastic models?

Estimation of GARCH Models Using
Maximum Likelihood
• Now we have yt =  + yt-1 + ut , ut  N(0, t2)

t2 = 0 + 1 ut21 +t-12
T 1 T 1 T
L   log( 2 )   log(  t )   ( yt    yt 1 ) 2 /  t
2 2
2 2 t 1 2 t 1
• Unfortunately, the LLF for a model with time-varying variances cannot be
maximised analytically, except in the simplest of cases. So a numerical
procedure is used to maximise the log-likelihood function. A potential
problem: local optima or multimodalities in the likelihood surface.
• The way we do the optimisation is:
1. Set up LLF.
2. Use regression to get initial guesses for the mean parameters.
3. Choose some initial guesses for the conditional variance parameters.
4. Specify a convergence criterion - either by criterion or by value.
Non-Normality and Maximum Likelihood
• Recall that the conditional normality assumption for ut is essential.
• We can test for normality using the following representation

ut = vtt vt  N(0,1)
ut
 t  0  1ut 1  2 t 1
2 2 v 
t
 t
uˆt
• The sample counterpart is vˆt 
ˆ t
• Are the v̂t normal? Typically v̂t are still leptokurtic, although less so than
the ût . Is this a problem? Not really, as we can use the ML with a robust
variance/covariance estimator. ML with robust standard errors is called Quasi-
Maximum Likelihood or QML.
Extensions to the Basic GARCH Model
• Since the GARCH model was developed, a huge number of extensions

and variants have been proposed. Three of the most important
examples are EGARCH, GJR, and GARCH-M models.
• Problems with GARCH(p,q) Models:

- Non-negativity constraints may still be violated
- GARCH models cannot account for leverage effects
• Possible solutions: the exponential GARCH (EGARCH) model or the

GJR model, which are asymmetric GARCH models.

The EGARCH Model
• Suggested by Nelson (1991). The variance equation is given by
u t 1  u 2
 
t 1
log(  t )     log(  t 1 )   
2 2
 t 1 2   t 1 2 
 
• Advantages of the model

- Since we model the log(t2), then even if the parameters are negative, t2
will be positive.
- We can account for the leverage effect: if the relationship between
volatility and returns is negative, , will be negative.

The GJR Model
• Due to Glosten, Jaganathan and Runkle

t2 = 0 + 1 ut21 +t-12+ut-12It-1
where It-1 = 1 if ut-1 < 0

= 0 otherwise
• For a leverage effect, we would see  > 0.
• We require 1 +   0 and 1  0 for non-negativity.

An Example of the use of a GJR Model
• Using monthly S&P 500 returns, December 1979- June 1998
• Estimating a GJR model, we obtain the following results.

y t  0.172
(3.198)
 t 2  1.243  0.015ut21  0.498 t 1 2  0.604ut21 I t 1

(16.372) (0.437) (14.999) (5.772)

News Impact Curves
The news impact curve plots the next period volatility (ht) that would arise from various
positive and negative values of ut-1, given an estimated model.
News Impact Curves for S&P 500 Returns using Coefficients from GARCH and GJR
Model Estimates: 0.14
GARCH
GJR
0.12
Value of Conditional Variance
0.1
0.08
0.06
0.04
0.02
0
-1 -0.9 -0.8 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Value of Lagged Shock
GARCH-in Mean
• We expect a risk to be compensated by a higher return. So why not let

the return of a security be partly determined by its risk?
• Engle, Lilien and Robins (1987) suggested the ARCH-M specification.

A GARCH-M model would be
yt =  + t-1+ ut , ut  N(0,t2)
t2 = 0 + 1 ut21 +t-12
•  can be interpreted as a sort of risk premium.
• It is possible to combine all or some of these models together to get

more complex “hybrid” models - e.g. an ARMA-EGARCH(1,1)-M
model.
What Use Are GARCH-type Models?
• GARCH can model the volatility clustering effect since the conditional
variance is autoregressive. Such models can be used to forecast volatility.
• We could show that

Var (yt  yt-1, yt-2, ...) = Var (ut  ut-1, ut-2, ...)
• So modelling t2 will give us models and forecasts for yt as well.
• Variance forecasts are additive over time.

Forecasting Variances using GARCH Models
• Producing conditional variance forecasts from GARCH models uses a

very similar approach to producing forecasts from ARMA models.
• It is again an exercise in iterating with the conditional expectations
operator.
• Consider the following GARCH(1,1) model:
yt    ut , ut  N(0,t2),  t 2   0   1u t21   t 1 2
• What is needed is to generate forecasts of T+12 T, T+22 T, ...,
T+s2 T where T denotes all information available up to and
including observation T.
• Adding one to each of the time subscripts of the above conditional
variance equation, and then two, and then three would yield the
following equations
T+12 = 0 + 1uT2 +T2 , T+22 = 0 + 1uT+12 +T+12 , T+32 = 0 + 1
uT+2+T+22
Forecasting Variances
using GARCH Models (Cont’d)
• Let 1f,T be the one step ahead forecast for 2 made at time T. This is
2
easy to calculate since, at time T, the values of all the terms on the
RHS are known.
•  1f,T 2 would be obtained by taking the conditional expectation of the
first equation at the bottom of slide 36:
 1f,T = 0 + 1 uT2 +T2
2
• Given,  1f,T how is 2f,T , the 2-step ahead forecast for 2 made at time T,
2 2
calculated? Taking the conditional expectation of the second equation

at the bottom of slide 36:
 2f,T = 0 + 1E( uT 1 T) +  1f,T
2 2 2
• where E( uT 1 T) is the expectation, made at time T, of uT 1, which is

2 2
the squared disturbance term.

Forecasting Variances
using GARCH Models (Cont’d)
• We can write
E(uT+12  t) = T+12
• But T+12 is not known at time T, so it is replaced with the forecast for
it,  f 2 , so that the 2-step ahead forecast is given by
1,T
 2f,T = 0 + 1 1f,T + 1f,T
2 2 2
f 2 =  + ( +) f
2
 2,T 0 1 1,T
• By similar arguments, the 3-step ahead forecast will be given by
 3f,T = ET(0 + 1uT+22 + T+22)
2
= 0 + (1+) 2f,T 2
= 0 + (1+)[ 0 + (1+) 1f,T ]
2
= 0 + 0(1+) + (1+)2 1f,T

2
• Any s-step ahead forecast (s  2) would be produced by

s 1
 f 2
  0  ( 1   ) i 1  ( 1   ) s 1  1f,T
2
s ,T
i 1
What Use Are Volatility Forecasts?
1. Option pricing
C = f(S, X, 2, T, rf)
2. Conditional betas
im,t
i ,t 
m2 ,t
3. Dynamic hedge ratios

The Hedge Ratio - the size of the futures position to the size of the
underlying exposure, i.e. the number of futures contracts to buy or sell per
unit of the spot good.
What Use Are Volatility Forecasts? (Cont’d)
• What is the optimal value of the hedge ratio?

• Assuming that the objective of hedging is to minimise the variance of the
hedged portfolio, the optimal hedge ratio will be given by
s
h p
F
where h = hedge ratio
p = correlation coefficient between change in spot price (S) and
change in futures price (F)
S = standard deviation of S
F = standard deviation of F
• What if the standard deviations and correlation are changing over time?
 s ,t
Use ht  p t
 F ,t
Testing Non-linear Restrictions or
Testing Hypotheses about Non-linear Models
• Usual t- and F-tests are still valid in non-linear models, but they are
not flexible enough.
• There are three hypothesis testing procedures based on maximum

likelihood principles: Wald, Likelihood Ratio, Lagrange Multiplier.
• Consider a single parameter,  to be estimated, Denote the MLE as ˆ

and a restricted estimate as ~ .

Likelihood Ratio Tests
• Estimate under the null hypothesis and under the alternative.

• Then compare the maximised values of the LLF.
• So we estimate the unconstrained model and achieve a given maximised
value of the LLF, denoted Lu
• Then estimate the model imposing the constraint(s) and get a new value of
the LLF denoted Lr.
• Which will be bigger?
• Lr  Lu comparable to RRSS  URSS
• The LR test statistic is given by

LR = -2(Lr - Lu)  2(m)
where m = number of restrictions

Likelihood Ratio Tests (cont’d)
• Example: We estimate a GARCH model and obtain a maximised LLF of

66.85. We are interested in testing whether  = 0 in the following equation.
yt =  + yt-1 + ut , ut  N(0, t )
2
t2 = 0 + 1 ut21 +   t 1
2
• We estimate the model imposing the restriction and observe the maximised
LLF falls to 64.54. Can we accept the restriction?
• LR = -2(64.54-66.85) = 4.62.
• The test follows a 2(1) = 3.84 at 5%, so reject the null.
• Denoting the maximised value of the LLF by unconstrained ML as L(ˆ)
~
and the constrained optimum as L( ) . Then we can illustrate the 3 testing
procedures in the following diagram:
Comparison of Testing Procedures under Maximum
Likelihood: Diagramatic Representation
L 
A

L ˆ
B
L
~
~
 ˆ 

Hypothesis Testing under Maximum Likelihood
• The vertical distance forms the basis of the LR test.
• The Wald test is based on a comparison of the horizontal distance.
• The LM test compares the slopes of the curve at A and B.
• We know at the unrestricted MLE, L(ˆ), the slope of the curve is zero.
~
• But is it “significantly steep” at L( ) ?
• This formulation of the test is usually easiest to estimate.

An Example of the Application of GARCH Models
- Day & Lewis (1992)
• Purpose
• To consider the out of sample forecasting performance of GARCH and
EGARCH Models for predicting stock index volatility.
• Implied volatility is the markets expectation of the “average” level of

volatility of an option:
• Which is better, GARCH or implied volatility?
• Data
• Weekly closing prices (Wednesday to Wednesday, and Friday to Friday)
for the S&P100 Index option and the underlying 11 March 83 - 31 Dec. 89
• Implied volatility is calculated using a non-linear iterative procedure.

The Models
• The “Base” Models

For the conditional mean
RMt  RFt  0  1 ht  ut (1)
And for the variance ht   0   1u t21   1 ht 1 (2)
ut 1  u  2 
1/ 2 
or ln( ht )   0  1 ln( ht 1 )  1 (    t 1    ) (3)
ht 1  ht 1    
where
RMt denotes the return on the market portfolio
RFt denotes the risk-free rate
ht denotes the conditional variance from the GARCH-type models while
t2 denotes the implied variance from option prices.
The Models (cont’d)
• Add in a lagged value of the implied volatility parameter to equations (2)

and (3).
(2) becomes
ht   0   1u t21   1 ht 1   t21 (4)
and (3) becomes

ut 1  u 2 
1/ 2
ln( ht )   0  1 ln( ht 1 )  1 (   t 1
   )   ln(  t21 ) (5)
ht 1  ht 1    
• We are interested in testing H0 :  = 0 in (4) or (5).

• Also, we want to test H0 : 1 = 0 and 1 = 0 in (4),
• and H0 : 1 = 0 and 1 = 0 and  = 0 and  = 0 in (5).
The Models (cont’d)
• If this second set of restrictions holds, then (4) & (5) collapse to
ht2   0   t21 (4’)
• and (3) becomes

ln( ht2 )   0   ln(  t21 ) (5’)
• We can test all of these restrictions using a likelihood ratio test.

In-sample Likelihood Ratio Test Results:
GARCH Versus Implied Volatility
R Mt  R Ft   0  1 ht  u t (8.78)
ht   0   1u t21   1 ht 1 (8.79)
ht   0   1u t21   1 ht 1   t21 (8.81)
ht2   0   t21 (8.81)
Equation for 0 1 010-4 1 1  Log-L 2
Variance
specification
(8.79) 0.0072 0.071 5.428 0.093 0.854 - 767.321 17.77
(0.005) (0.01) (1.65) (0.84) (8.17)
(8.81) 0.0015 0.043 2.065 0.266 -0.068 0.318 776.204 -
(0.028) (0.02) (2.98) (1.17) (-0.59) (3.00)
(8.81) 0.0056 -0.184 0.993 - - 0.581 764.394 23.62
(0.001) (-0.001) (1.50) (2.94)
Notes: t-ratios in parentheses, Log-L denotes the maximised value of the log-likelihood function in
each case. 2 denotes the value of the test statistic, which follows a 2(1) in the case of (8.81) restricted
to (8.79), and a 2 (2) in the case of (8.81) restricted to (8.81). Source: Day and Lewis (1992).
Reprinted with the permission of Elsevier Science.
In-sample Likelihood Ratio Test Results:
EGARCH Versus Implied Volatility
R Mt  R Ft   0  1 ht  u t (8.78)
u t 1  u t 1 2
1/ 2
ln( ht )   0   1 ln( ht 1 )   1 (     ) (8.80)
ht 1  ht 1   
u t 1  u t 1 2 
1/ 2
ln( ht )   0   1 ln( ht 1 )   1 (       )   ln(  t21 ) (8.82)

ht 1  ht 1    
ln( ht2 )   0   ln(  t21 ) (8.82)
Equation for 0 1 010 -4
1    Log-L 2
Variance
specification
(c) -0.0026 0.094 -3.62 0.529 -0.273 0.357 - 776.436 8.09
(-0.03) (0.25) (-2.90) (3.26) (-4.13) (3.17)
(e) 0.0035 -0.076 -2.28 0.373 -0.282 0.210 0.351 780.480 -
(0.56) (-0.24) (-1.82) (1.48) (-4.34) (1.89) (1.82)
(e) 0.0047 -0.139 -2.76 - - - 0.667 765.034 30.89
(0.71) (-0.43) (-2.30) (4.01)
Notes: t-ratios in parentheses, Log-L denotes the maximised value of the log-likelihood function in
each case. 2 denotes the value of the test statistic, which follows a 2(1) in the case of (8.82) restricted
to (8.80), and a 2 (2) in the case of (8.82) restricted to (8.82). Source: Day and Lewis (1992).

Conclusions for In-sample Model Comparisons &
Out-of-Sample Procedure
• IV has extra incremental power for modelling stock volatility beyond

GARCH.
• But the models do not represent a true test of the predictive ability of
IV.
• So the authors conduct an out of sample forecasting test.
• There are 729 data points. They use the first 410 to estimate the
models, and then make a 1-step ahead forecast of the following week’s
volatility.
• Then they roll the sample forward one observation at a time,

constructing a new one step ahead forecast at each step.
Out-of-Sample Forecast Evaluation
• They evaluate the forecasts in two ways:

• The first is by regressing the realised volatility series on the forecasts plus
a constant:
t21  b0  b1 2ft  t 1 (7)
where t 1 is the “actual” value of volatility, and  2ft is the value forecasted
2
for it during period t.

• Perfectly accurate forecasts imply b0 = 0 and b1 = 1.
• But what is the “true” value of volatility at time t ?
Day & Lewis use 2 measures
1. The square of the weekly return on the index, which they call SR.
2. The variance of the week’s daily returns multiplied by the number
of trading days in that week.

Out-of Sample Model Comparisons
 t21  b0  b1 2ft   t 1 (8.83)

Forecasting Model Proxy for ex b0 b1 R2
post volatility
Historic SR 0.0004 0.129 0.094
(5.60) (21.18)
Historic WV 0.0005 0.154 0.024
(2.90) (7.58)
GARCH SR 0.0002 0.671 0.039
(1.02) (2.10)
GARCH WV 0.0002 1.074 0.018
(1.07) (3.34)
EGARCH SR 0.0000 1.075 0.022
(0.05) (2.06)
EGARCH WV -0.0001 1.529 0.008
(-0.48) (2.58)
Implied Volatility SR 0.0022 0.357 0.037
(2.22) (1.82)
Implied Volatility WV 0.0005 0.718 0.026
(0.389) (1.95)
Notes: Historic refers to the use of a simple historical average of the squared returns to forecast
volatility; t-ratios in parentheses; SR and WV refer to the square of the weekly return on the S&P 100,
and the variance of the week’s daily returns multiplied by the number of trading days in that week,
respectively. Source: Day and Lewis (1992). Reprinted with the permission of Elsevier Science.

Encompassing Test Results: Do the IV Forecasts
Encompass those of the GARCH Models?
 t21  b0  b1 It2  b2 Gt2  b3 Et2  b4 Ht
2
  t 1 (8.86)
Forecast comparison b0 b1 b2 b3 b4 R2
-0.00010 0.601 0.298 - - 0.027
Implied vs. GARCH (-0.09) (1.03) (0.42)
Implied vs. GARCH 0.00018 0.632 -0.243 - 0.123 0.038

vs. Historical (1.15) (1.02) (-0.28) (7.01)
Implied vs. EGARCH -0.00001 0.695 - 0.176 - 0.026

(-0.07) (1.62) (0.27)
Implied vs. EGARCH 0.00026 0.590 -0.374 - 0.118 0.038

vs. Historical (1.37) (1.45) (-0.57) (7.74)
GARCH vs. EGARCH 0.00005 - 1.070 -0.001 - 0.018

(0.37) (2.78) (-0.00)
Notes: t-ratios in parentheses; the ex post measure used in this table is the variance of the week’s daily
returns multiplied by the number of trading days in that week. Source: Day and Lewis (1992).
Conclusions of Paper
• Within sample results suggest that IV contains extra information not

contained in the GARCH / EGARCH specifications.
• Out of sample results suggest that nothing can accurately predict

volatility!

Stochastic Volatility Models
• It is a common misconception that GARCH-type specifications are

stochastic volatility models
• However, as the name suggests, stochastic volatility models differ
from GARCH principally in that the conditional variance equation of a
GARCH specification is completely deterministic given all
information available up to that of the previous period
• There is no error term in the variance equation of a GARCH model,
only in the mean equation
• Stochastic volatility models contain a second error term, which enters
into the conditional variance equation.

Autoregressive Volatility Models
• A simple example of a stochastic volatility model is the autoregressive

volatility specification
• This model is simple to understand and simple to estimate, because it
requires that we have an observable measure of volatility which is then
simply used as any other variable in an autoregressive model
• The standard Box-Jenkins-type procedures for estimating
autoregressive (or ARMA) models can then be applied to this series
• For example, if the quantity of interest is a daily volatility estimate, we
could use squared daily returns, which trivially involves taking a
column of observed returns and squaring each observation
• The model estimated for volatility, t2, is then

A Stochastic Volatility Model Specification
• The term ‘stochastic volatility’ is usually associated with a different

formulation to the autoregressive volatility model, a possible example
of which would be
where ηt is another N(0,1) random variable that is independent of ut

• The volatility is latent rather than observed, and so is modelled
indirectly
• Stochastic volatility models are superior in theory compared with
GARCH-type models, but the former are much more complex to
estimate.

Covariance Modelling: Motivation
• A limitation of univariate volatility models is that the fitted conditional

variance of each series is entirely independent of all others
• This is potentially an important limitation for two reasons:
– If there are ‘volatility spillovers’ between markets or assets, the univariate
model will be mis-specified
– It is often the case that the covariances between series are of interest too
– The calculation of hedge ratios, portfolio value at risk estimates, CAPM
betas, and so on, all require covariances as inputs
• Multivariate GARCH models can be used for estimation of:
– Conditional CAPM betas
– Dynamic hedge ratios
– Portfolio variances

Simple Covariance Models:
Historical and Implied
• In exactly the same fashion as for volatility, the historical covariance

or correlation between two series can be calculated from a set of
historical data
• Implied covariances can be calculated using options whose payoffs are

dependent on more than one underlying asset
• The relatively small number of such options that exist limits the
circumstances in which implied covariances can be calculated
• Examples include rainbow options, ‘crack spread’ options for different

grades of oil, and currency options.

Implied Covariance Models
• To give an illustration for currency options, the implied variance of the

cross-currency returns is given by
where and are the implied variances of the x and y

returns respectively, and is the implied covariance between
x and y
• So if the implied covariance between USD/DEM and USD/JPY is of

interest, then the implied variances of the returns of USD/DEM and
USD/JPY and the returns of the cross-currency DEM/JPY are required.

EWMA Covariance Models
• A EWMA specification gives more weight in the covariance to recent

observations than an estimate based on the simple average
• The EWMA model estimates for variances and covariances at time t in
the bivariate setup with two returns series x and y may be written as
hij,t = λhij,t−1 + (1 − λ)xt−1yt−1
where i  j for the covariances and i = j; x = y for the variances
• The fitted values for h also become the forecasts for subsequent
periods
• λ (0 < λ < 1) denotes the decay factor determining the relative weights
attached to recent versus less recent observations
• This parameter could be estimated but is often set arbitrarily (e.g.,
Riskmetrics use a decay factor of 0.97 for monthly data but 0.94 for
daily).
EWMA Covariance Models - Limitations
• This equation can be rewritten as an infinite order function of only the

returns by successively substituting out the covariances:
• The EWMA model is a restricted version of an integrated GARCH

(IGARCH) specification, and it does not guarantee the fitted variance-
covariance matrix to be positive definite
• EWMA models also cannot allow for the observed mean reversion in
the volatilities or covariances of asset returns that is particularly
prevalent at lower frequencies.

Multivariate GARCH Models
• Multivariate GARCH models are used to estimate and to forecast

covariances and correlations.
• The basic formulation is similar to that of the GARCH model, but where the
covariances as well as the variances are permitted to be time-varying.
• There are 3 main classes of multivariate GARCH formulation that are widely
used: VECH, diagonal VECH and BEKK.
VECH and Diagonal VECH

• e.g. suppose that there are two variables used in the model. The conditional
covariance matrix is denoted Ht, and would be 2  2. Ht and VECH(Ht) are
 h11t 
 h11t h12t 
Ht    VECH ( H t )  h22t 
h21t h22t   h12t 
VECH and Diagonal VECH
• In the case of the VECH, the conditional variances and covariances would
each depend upon lagged values of all of the variances and covariances
and on lags of the squares of both error terms and their cross products.
• In matrix form, it would be written
VECH H t   C  A VECH  t 1t 1   B VECH H t 1   t  t 1 ~ N 0, H t 
• Writing out all of the elements gives the 3 equations as
h11t  c11  a11u12t  a12 u 22t  a13 u1t u 2t  b11 h11t 1  b12 h22 t 1  b13 h12 t 1
h22 t  c21  a 21u12t  a 22 u 22t  a 23 u1t u 2 t  b21 h11t 1  b22 h22 t 1  b23 h12 t 1
h12 t  c31  a 31u12t  a 32 u 22t  a 33 u1t u 2 t  b31 h11t 1  b32 h22 t 1  b33 h12 t 1
• Such a model would be hard to estimate. The diagonal VECH is much
simpler and is specified, in the 2 variable case, as follows:
h11t   0   1u12t 1   2 h11t 1
h22t   0   1u 22t 1   2 h22t 1
h12t   0   1u1t 1u 2t 1   2 h12t 1
BEKK and Model Estimation for M-GARCH
• Neither the VECH nor the diagonal VECH ensure a positive definite
variance-covariance matrix.
• An alternative approach is the BEKK model (Engle & Kroner, 1995).
• The BEKK Model uses a Quadratic form for the parameter matrices to
ensure a positive definite variance / covariance matrix Ht.
• In matrix form, the BEKK model is Ht  W W  AHt 1 A  Bt 1t 1B
• Model estimation for all classes of multivariate GARCH model is
again performed using maximum likelihood with the following LLF:
   
TN
2
1 T

log 2   log H t   t' H t1 t
2 t 1

where N is the number of variables in the system (assumed 2 above), 
is a vector containing all of the parameters, and T is the number of obs.

Correlation Models and the CCC
• The correlations between a pair of series at each point in time can be

constructed by dividing the conditional covariances by the product of
the conditional standard deviations from a VECH or BEKK model
• A subtly different approach would be to model the dynamics for the
correlations directly
• In the constant conditional correlation (CCC) model, the correlations
between the disturbances to be fixed through time
• Thus, although the conditional covariances are not fixed, they are tied
to the variances
• The conditional variances in the fixed correlation model are identical
to those of a set of univariate GARCH specifications (although they
are estimated jointly):

More on the CCC
• The off-diagonal elements of Ht, hij,t (i  j), are defined indirectly via
the correlations, denoted ρij:
• Is it empirically plausible to assume that the correlations are constant

through time?
• Several tests of this assumption have been developed, including a test

based on the information matrix due and a Lagrange Multiplier test
• There is evidence against constant correlations, particularly in the

context of stock returns.
The Dynamic Conditional Correlation Model
• Several different formulations of the dynamic conditional correlation

(DCC) model are available, but a popular specification is due to Engle
(2002)
• The model is related to the CCC formulation but where the correlations
are allowed to vary over time.
• Define the variance-covariance matrix, Ht, as Ht = DtRtDt
• Dt is a diagonal matrix containing the conditional standard deviations
(i.e. the square roots of the conditional variances from univariate
GARCH model estimations on each of the N individual series) on the
leading diagonal
• Rt is the conditional correlation matrix
• Numerous parameterisations of Rt are possible, including an
exponential smoothing approach
The DCC Model – A Possible Specification
• A possible specification is of the MGARCH form:
where:
• S is the unconditional correlation matrix of the vector of standardised
residuals (from the first stage estimation), ut = Dt−1ϵt
• ι is a vector of ones
• Qt is an N × N symmetric positive definite variance-covariance matrix
• ◦ denotes the Hadamard or element-by-element matrix multiplication
procedure
• This specification for the intercept term simplifies estimation and
reduces the number of parameters.
The DCC Model – A Possible Specification
• Engle (2002) proposes a GARCH-esque formulation for dynamically

modelling Qt with the conditional correlation matrix, Rt then constructed
as
where diag(·) denotes a matrix comprising the main diagonal elements

of (·) and Q∗ is a matrix that takes the square roots of each element in Q
• This operation is effectively taking the covariances in Qt and dividing

them by the product of the appropriate standard deviations in Qt∗ to
create a matrix of correlations.

DCC Model Estimation
• The model may be estimated in a single stage using ML although this will
be difficult. So Engle advocates a two-stage procedure where each variable
in the system is first modelled separately as a univariate GARCH
• A joint log-likelihood function for this stage could be constructed, which
would simply be the sum (over N) of all of the log-likelihoods for the
individual GARCH models
• In the second stage, the conditional likelihood is maximised with respect
to any unknown parameters in the correlation matrix
• The log-likelihood function for the second stage estimation will be of the
form
• where θ1 and θ2 denote the parameters to be estimated in the 1st and 2nd
stages respectively.
Asymmetric Multivariate GARCH
• Asymmetric models have become very popular in empirical applications,

where the conditional variances and / or covariances are permitted to react
differently to positive and negative innovations of the same magnitude
• In the multivariate context, this is usually achieved in the Glosten et al.
(1993) framework
• Kroner and Ng (1998), for example, suggest the following extension to the
BEKK formulation (with obvious related modifications for the VECH or
diagonal VECH models)
where zt−1 is an N-dimensional column vector with elements taking the

value −ϵt−1 if the corresponding element of ϵt−1 is negative and zero
otherwise.

An Example: Estimating a Time-Varying Hedge Ratio
for FTSE Stock Index Returns
(Brooks, Henry and Persand, 2002).
• Data comprises 3580 daily observations on the FTSE 100 stock index and
stock index futures contract spanning the period 1 January 1985 - 9 April 1999.
• Several competing models for determining the optimal hedge ratio are
constructed. Define the hedge ratio as .
– No hedge (=0)
– Naïve hedge (=1)
– Multivariate GARCH hedges:
• Symmetric BEKK
• Asymmetric BEKK
In both cases, estimating the OHR involves forming a 1-step ahead
forecast and computing
hCF ,t 1
OHR t 1   t
hF ,t 1
OHR Results
In Sample
Unhedged Naïve Hedge Symmetric Time Asymmetric
=0 =1 Varying Time Varying
Hedge Hedge
hFC ,t hFC ,t
t  t 
h F ,t h F ,t
Return 0.0389 -0.0003 0.0061 0.0060
{2.3713} {-0.0351} {0.9562} {0.9580}
Variance 0.8286 0.1718 0.1240 0.1211
Out of Sample
Unhedged Naïve Hedge Symmetric Time Asymmetric
=0 =1 Varying Time Varying
Hedge Hedge
hFC ,t hFC ,t
t  t 
h F ,t hF ,t
Return 0.0819 -0.0004 0.0120 0.0140
{1.4958} {0.0216} {0.7761} {0.9083}
Variance 1.4972 0.1696 0.1186 0.1188

Plot of the OHR from Multivariate GARCH
Time Varying Hedge Ratios

1.00
0.95
Conclusions
0.90
- OHR is time-varying and less
than 1
0.85 - M-GARCH OHR provides a
better hedge, both in-sample
0.80
and out-of-sample.
0.75
- No role in calculating OHR for
asymmetries
0.70
0.65
500 1000 1500 2000 2500 3000
Symmetric BEKK
Asymmetric BEKK

Modelling Volatility and Correlation

Uploaded by

Copyright:

Available Formats

Modelling Volatility and Correlation

Uploaded by

Document Information

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

Modelling Volatility and Correlation

Uploaded by

Copyright:

Available Formats

Chapter 9

Modelling volatility and correlation

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 1

• Motivation: the linear structural (and time series) models cannot

• Our “traditional” structural model could be something like:

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 2

Daily S&P 500 Returns for August 2003 – August 2013

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 3

• Campbell, Lo and MacKinlay (1997) define a non-linear data

• They also give a slightly more specific definition as

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 4

• The linear paradigm is a useful one.

• There are many types of non-linear models, e.g.

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 5

• The “traditional” tools of time series analysis (acf’s, spectral analysis)

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 6

• Chaos theory is a notion taken from the physical sciences

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 8

• A system is chaotic if it exhibits sensitive dependence on initial

• Artificial neural networks (ANNs) are a class of models whose

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 10

• The most common class of ANN models in finance are known as

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 11

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 12

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 13

• The simplest model for volatility is the historical estimate

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 14

• An example of a structural model is

• The assumption that the variance of the errors is constant is known as

• What if the variance of the errors is not constant?

We usually assume that E(ut) = 0

• This is an ARCH(q) model.

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 17

• For illustration, consider an ARCH(1). Instead of the above, we can

yt = 1 + 2x2t + ... + kxkt + ut , ut = vtt

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 18

3. The test statistic is defined as TR2 (the number of observations multiplied

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 19

4. The null and alternative hypotheses are

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 20

• A natural extension of an ARCH(q) model which gets around some of

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 21

• Due to Bollerslev (1986). Allow the conditional variance to be dependent

• Now substituting into (2) for t-22

• We can again extend the GARCH(1,1) model to a GARCH(p,q):

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 23

• But in general a GARCH(1,1) model will be sufficient to capture the

• Why is GARCH Better than ARCH?

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 24

• The unconditional variance of ut is given by

• 1    1 is termed “non-stationarity” in variance

• 1   = 1 is termed intergrated GARCH

• For non-stationarity in variance, the conditional variance forecasts will

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 25

• We use another technique known as maximum likelihood.

• More specifically, we form a log-likelihood function and maximise it.

‘Introductory Econometrics for Finance’ © Chris Brooks 2013 26

• The steps involved in actually estimating an ARCH or GARCH model

• Consider the bivariate regression case with homoscedastic errors for

• Assuming that ut  N(0,2), then yt  N( 1   2 xt , 2) so that the