Applied Macro and Financial

Econometrics

Volatility Clustering

Course Coordinator:
Dr. Devasmita Jena
Volatility Clustering: An Introduction
Many economic TS (macro/financial variables) exhibit period of unusually large volatility
followed by period of relative tranquility
There are periods when a series shows higher or lower variance than other time periods
You will observe clustering: groups of high variance followed by groups of low variance
This is called volatility clustering => series exhibits time varying heteroscedasticity
That is, for a large class of models, the size of volatility is not constant and varies with time
In econometric sense, volatility measures the size of errors made in modelling TS variables
Varying volatility is predictable and hence can be modelled using appropriate methodology
Example: As a firm, you will be interested in both mean and variance (and may be entire
distribution) of the investments. Why?
• You face trade-off between returns and risks
• Assume a large shock to return at t-1, after which there is high probability of shock at t.
• The shock has upset the market and a large uncertainty on the future direction of returns follow
• This is volatility clustering : a reflection of high and low market uncertainty
Do not confuse between volatility and NS: volatility is time-varying but not function of time
Volatility Clustering: Visual Examples
Modelling Volatility Clustering: Intuition
Suppose {Yt} is the TS representing returns from at asset; heteroscedasticity can be modeled
as:
V(Yt|Xt) = ; where {Xt} is any variable that influences the variance of {Yt} to change over
time.
Question is: what is Xt (in this example of returns)?
• Inflation rate (due to change in monetary policy, perhaps)
• Oil shock
It is sometimes difficult to identify {Xt} series; also there could be multiple {Xt} that can
influence the variance of {Yt} to change, conditional on the change in {Xt}
Method needs to “account” for the time varying heteroscedastcity
Note: variance of {Yt}changes conditional on another series which may or may not be
known to us. But unconditional variance has to remain the same. Why?
• NS condition
• Volatility clustering ~ volatility is autocorrelated
What happens if you ignore volatility clustering?
What happens if we ignore conditional variance of {Yt}with time and that there is
volatility clustering?
• The standard error of forecasts of Yt will be large; substantially different for different
time periods
It is sometimes of interest to not only forecast the mean of the series but also its
variance over the study period
ARCH, GARCH, GJR-GARCH, TGARCH, EGARCH etc.
Example: Suppose you buy the asset at t and want to sell it at t+1.
• In this case, we’re interested in forecasting the conditional variance of Yt+1 given the
variance is known at t (or t-1, t-2, t-3,...)
• The unconditional variance which is the long run forecast variance of the series is not of
importance. Why?
• Because unconditional variance has to be constant and not a function of time!
Unconditional variance
provides a baseline or
equilibrium level to
which the volatility
reverts in the long run
What happens if you ignore volatility clustering?

This graph represents the forecasts

by 3 different models and the actual
values (represented by violet line).
ARIMA (4,1,0) (the red line) which
has the lowest BIC value seems to
perform better than other models.
Autoregressive conditionally heteroscedastic models
 The build up of ARCH model comes from the definition of conditional variance of and the distinction
between the conditional and unconditional variances of
 The conditional variance of may be denoted , which is written as:
]
=> conditional variance of a zero mean normally distributed random variable is equal to the conditional
expected value of the square of
 Under the ARCH model, the ‘volatility clustering’ is modelled by allowing the conditional variance of the
error term, , to depend on the immediately previous value of the squared error:

 This is an ARCH(1), since the conditional variance depends on only one lagged squared error
 Full model:
; ~N(0,

This is AR(1)-ARCH(1)model
 Generalized ARCH(q) model:
Non-Negativity constraint in ARCH models
Since is a conditional variance, its value must always be strictly positive; a negative
variance at any point in time would be meaningless
The variables on the RHS of the conditional variance equation are all squares of lagged
errors, and so can’t be negative.
To ensure that positivity of , all of the coefficients in the conditional variance should be non-
negative
If one or more of the coefficients were to take on a negative value, then for a sufficiently
large lagged squared error term attached to that coefficient, the fitted value from the model
for the conditional variance could be negative
This wont make sense
Therefore, for an ARCH(q) model, αi ≥ 0 ∀ i = 0, 1, 2,..., q to ensure to be positive
This is a slightly stronger condition than is actually necessary for non-negativity of the
conditional variance
Testing for ARCH Effects
The test is basically a test for autocorrelation in the squared residuals
ARCH test is applied to both the residuals of an estimated model and raw data
Steps:
1. Run ARMA model, and get the estimated residuals

2. Square the residuals, and regress them on m own lags to test for ARCH of order m:

3. Obtain R2 from this regression of residuals

4. The test statistic is defined as TR 2 (the number of observations multiplied by the coefficient of
multiple correlation) from the regression in step 2, and is distributed as a χ 2(m)
5. Hypotheses to test:
• H0 : γ1 = 0 and γ2 = 0 and γ3 = 0 and ... and γm = 0
• H1 : γ1 ≠ 0 or γ2 ≠ 0 or γ3 ≠ 0 or ... or γm ≠ 0
6. Value of the test statistic >critical value of distribution χ 2 => reject null hypothesis => ARCH
effects => volatility clustering is present
Limitations of ARCH Model
Deciding lag of ARCH
model
Lag of ARCH model and
parsimony of the model
Non-negativity constraint
may be violated
Generalized ARCH (GARCH) Model
 Developed by Bollerslev(1986) and Taylor(1986)
GARCH model allows the conditional variance to depend on pervious own lags as well:
; GARCH (1,1) model --- (Eq. 1)
As usual, is the conditional variance since it is a 1-pd ahead estimate for the variance
calculated based on any relevant past information
Current fitted variance, , can be interpreted as a weighted function of:
• A long-term average value ()
• Information about volatility during the previous period ()
• Fitted variance from the model during the previous period ()
GARCH model can be expressed as an ARMA model for conditional variance
Suppose you want to model . Then, Now, = - . Substituting this in Eq. 1:

This is expression for ARMA(1,1) process for squared errors

GARCH is parsimonious
 GARCH is parsimonious and avoids overfitting => less likely to violate non-negativity
constraint
Consider GARCH model

Taking infinite number of iterative substitution, we get:

; so that 1st expression of RHS: constant and 3rd expression is zero. We, then , have:
; where
This is restricted infinite order ARCH model
GARCH model requires estimation of only 3 parameters (, and ), making it parsimonious
GARCH allows infinite number of past squared errors to influence the current conditional
variance
Lag lengths of GARCH model
 GARCH(1,1) model can be extended to GARCH(p,q) model
GARCH(p,q) model: current conditional variance is dependent on q lags of squared error and p
lags of the conditional variance:

How to decide the lags p,q of GARCH(p,q) model?

It is a theoretical question and complicated at that!
Saving grace: A GARCH(1,1) model is sufficient to capture the volatility clustering in the data
and rarely is any higher order model estimated
The unconditional variance under GARCH
 Taking expectation of :

While the conditional variance is changing, the unconditional variance of is constant, so long
as
Some jargons!
• Non-stationarity in variance:
• Integrated GARCH/ IGARCH/ unit root in variance :
For stationary GARCH models, conditional variance forecasts convergence upon the long term
average value of the variance as the prediction horizon increases
For IGARCH process, this convergence doesn’t happen
For , the conditional variance forecast will tend to infinity as the forecast horizon increases!
Estimation of Models with Volatility Clusters
 OLS doesn’t work
• Usual suspect: Autocorrelation
• Full model is no longer of the usual linear form
• OLS minimizes the residual sum of squares.
o The RSS depends only on the parameters in the conditional mean equation, and not the conditional variance
o RSS minimization is no longer an appropriate objective.

So, resort to MLE

Steps
• Specify appropriate model
• Specify log-likelihood function to maximize under normality assumption of disturbances
o Function taking into account the sample size (time horizon), the mean equation and variance equation
• Maximize LLF to obtain the estimated parameters along with standard errors
Issues with GARCH
 Non-negativity conditions may be violated by the estimated model
• place artificial constraints on the model coefficients in order to force them to be non-negative
Account for volatility clustering may not be enough, in the presence of leverage effect. GARCH
models cannot account for leverage effects
• GARCH enforce a symmetric response of volatility to +ve and –ve shocks
o Conditional variance is a function of the magnitudes of the lagged residuals and not their signs(squaring of lagged
errors leads to loss of sign/direction)
• -ve shock to macro-financial time series is likely to cause a greater extent of volatility clustering as compare
to a –ve shock of the same magnitude
• Good news (+ve shock) has less effect on conditional variance as compared to bad news (-ve shock)
GARCH cannot account for leptokurtosis in a series
• Many financial series, such as returns on stocks and foreign exchange rates, exhibit leptokurtosis
• GARCH model assume normality
GARCH does not allow for any direct feedback between the conditional variance and the
conditional mean
GJR-GARCH Model
 Due to Glosten, Jagannathan, Runkle(1993)
GJR model is a simple extension of GARCH with an additional term added to account for
possible asymmetries:

In the presence of leverage effect: γ > 0

Condition for non-negativity: α0 > 0, α1 > 0, β ≥ 0, and α1 + γ ≥ 0
γ < 0 is ok, provided that α1 + γ ≥ 0
Exponential-GARCH Model
 Due to Nelson(1991)
EGARCH can be expressed in various ways
One possible specification:
]
Advantages of EGARCH over GARCH:
• Since the log() is modelled, then even if the parameters are negative, will be positive
o no need to artificially impose non-negativity constraints on the model parameters
• Asymmetries are allowed under the EGARCH formulation
o if the relationship between volatility and returns is negative, γ , will be negative
o -ve shock generates higher volatility
Choosing order of GARCH model
 The ARMA and GARCH orders need to be determined simultaneously
 Neither the conditional mean nor the conditional variance model can be estimated consistently if taken
separately, unless under special condition: if MA part of ARMA is empty, the AR part can be estimated
consistently even if the GARCH model is neglected)
The task of jointly determining the ARMA-GARCH orders is difficult.
Idea is to find out order of autocorrelation in conditional variance. This will determine lag order
Following steps may be taken:
 Check info criteria of the system as a whole
 Estimate model on part sample and use the estimated model to predict the remainder of the smaple
 Pick the model that has lowest prediction error
 Examine the residual properties of different models
 Prefer a model with no remaining autocorrelation or ARCH patterns
 Check likelihoods of the model
All the above steps amount to trials and error and not full proof
The good news is: GARCH(1,1) usually suffices