MODELLING

VOLATILITY AND
CORELLATION
An introduction to News Impact Curve

Trizar Rizqiawan
Intro to non-linearity world
• We are familiar with linear models such as:

• The model above compactly written as Y=XB+u, assuming that

u follows normality with σ2
• Many relationships in finance are intrinsically non-linear, for
example: investors’ willingness to trade off returns and risks are
non-linear
• Some features that financial data tend to have are:
• Leptokurtosis: tendency for financial asset returns to have distributions
that exhibit fat tails and excess peakedness at the mean.
• Volatility clustering or volatility pooling: large returns follows large
returns and small returns to follow small returns.
• Leverage effects: tendency for volatility to rise more following a large
price fall than following a price rise of the same magnitude.
The initial ideas of non-linearity
• Current value of series is related non-linearly to current
and previous values of error term:

• Model with non-linear in mean and non-linear in variance:

• Model can be linear in mean and variance (eg: CLRM,

ARMA) or linear in mean but non-linear in variance (eg:
GARCH)
• Models can also be non-linear in mean but linear in
variance (bicorrelations models)
Testing for non-linearity
• A non-linear model should be used where Tests for non-
linear
financial theory suggests that the
relationship between variables should be General test
(portmanteau)
Specific test

such as to require a non-linear model.

Ramsey’s
• BDS and Ramsey RESET tests are RESET Test

available in E-views
BDS test

Bispectrum
test

Bicorrelation
test
Volatility
• Volatility is measured by the standard deviation or
variance of returns, often used as a crude measure of the
total risk of financial assets.
• Volatility is usually calculated as the variance or standard
deviation of returns over some historical periods.
Modelling Volatility - EWMA
• Exponentially Weighted Moving Average
• The idea is that recent observations have a stronger
impact on the forecast of volatility than older data.
• Greater weight is put on recent data.
• The effect of older data is decreasing exponentially.
• The model:

• Limitations: need correction if we only have small samples

Modelling Volatility - Autoregressive
volatility model
• The idea is that a time series of observations on some
volatility proxy are obtained, we can use Box-Jenkins
procedures to estimate the model.
• Daily volatility can be calculated as :

• We can estimate the model using ARMA:

Modelling Volatility - ARCH
• Autoregressive conditionally heteroscedastic
• The assumptions are similar to linear regressions
although variance is not assumed to be constant in
ARCH.
• ARCH model accommodates “volatility clustering” or
“volatility pooling”
Limitations of ARCH(q)
• How we decide the value of q (the number of lags), can
use likelihood but doesn’t guarantee best result
• Value of q, the number of lags of squared error that are
required to capture all the dependence in the conditional
variance, might be very large. We can use declining
length such as ARCH(4):

• Non-negativity constrains might be violated.

Modelling volatility - GARCH
• Generalised ARCH model
• It allows the conditional variance to be dependent upon
previous own lags : GARCH (1,1)

• Thus the model not only using the volatility of the basic
model but also the fitted variance from the model during
the previous period.
Tests for asymmetries in volatility
• Two popular asymmetric formulations are the GJR model
(also known as TGARCH) and EGARCH.
• Bad news tend to bring higher volatility than good news.
In other words, negative shocks tend to cause volatility
rise by more than positive shocks.
• We can test asymmetries using Engle and Ng (1993) test,
you can find the mechanics of the test on Brooks page
444.
News impact curve
• Pictorial representation of the degree of asymmetry of
volatility to positive and negative shocks (Pagan and
Schwert, 1990).
• The news impact curve plots the next-period volatility (σ2)
that would arise from various positive and negative values
of ut-1, given estimated model.