ARCH & GARCH Model

Introduction
• Financial time series, such as stock prices, interest rates, foreign exchange rates,
and inflation rates, often exhibit the phenomenon of volatility clustering.
• That is, periods of turbulence in which their prices show wide swings and
periods of tranquility in which there is relative calm.
• An average investor is not only concerned about the rate of return on his or her
investment, but also about the risk of investment as well as the variability, or
volatility, of risk.
• It is, therefore, important to measure asset price and asset returns volatility.
• A simple measure of asset return volatility is its variance over time.

• If we have data for stock returns over, say, a period of 1,000 days, we can
compute the variance of daily stock returns by subtracting the mean value of
stock returns from their individual values, square the difference and divide it by
the number of observations.

• By itself it does not capture volatility clustering because it is a measure of what

is called unconditional variance, which is a single number for a given sample.
• It does not take into account the past history of returns. That is, it does not take
into account time-varying volatility in asset returns.
• A measure that takes into account the past history is known as autoregressive
conditional heteroscedasticity, or ARCH for short
ARCH Model
• We usually encounter heteroscedasticity, or unequal variance, in cross-sectional
data because of the heterogeneity among individual cross-section units that
comprise cross-sectional observations, such as families, firms, regions, and
countries.
• We also usually observe autocorrelation in time series data.
• But in time series data involving asset returns, such as returns on stocks or foreign
exchange, we observe autocorrelated heteroscedasticity.
• That is, heteroscedasticity observed over different periods is autocorrelated.
• In the literature such a phenomenon is called autoregressive conditional
heteroscedasticity (ARCH).
• To get a glimpse of the daily dollar/euro exchange rate (EX), Figure 15.1 plots
the log of EX (LEX) for the sample period.
• Initially the EU was depreciating against the dollar, but later it showed a steady
appreciation against the dollar.
• But a closer look at the figure suggests that the initial depreciation and then
appreciation of EU was not smooth, which is apparent from the jagged nature
of the graph.
• This would suggest that there is considerable volatility of the dollar/euro
exchange rate.
• This can be seen more vividly if we plot the changes in the LEX (Figure 15.2); as
noted, changes in the log values represent relative changes, or percentage
changes if multiplied by 100.
• If you draw a horizontal line through 0.00, you will see clearly the volatility of
log-exchange rate changes: the amplitude of the change swings wildly from
time to time.
• Not only that, it seems there is a persistence in the swings that lasts for
sometime. That is, these swings seem to be autocorrelated. This is the heuristic
idea behind ARCH.
• A simple way to measure the volatility is to run the following regression:

• where RET is daily return and where c is a constant and ut represents the error term.
• Here we measure return as log changes in the exchange rate over successive days.
• The constant c in this case measures simply the mean value of daily exchange rate
returns.
• Notice that we have not introduced any explanatory variables in Eq. (15.1), for asset
returns are essentially unpredictable.
• if you obtain residuals from this regression (et) (which are simply the deviations of
daily returns from their mean value) and square them, you get the plot in Figure 15.3.
• This shows wide swings in the squared residuals, which can be taken as an
indicator of underlying volatility in the foreign exchange returns.
• Observe that not only are there clusters of periods when volatility is high and
clusters of periods when volatility is low, but these clusters seem to be
“autocorrelated”.
• That is, when volatility is high, it continues to be high for quite some time and
when volatility is low, it continues to be low for a while.
• How do we measure this volatility?
• The ARCH model attempt to answer this question.
• Consider the following simple linear regression model:

• This states that, conditional on the information available up to time (t – 1), the
value of the random variable Yt (exchange rate return here) is a function of the
variable Xt (or a vector of variables if there are more Xt variables) and ut.

• That is, given the information available up to time (t – 1), the error term is
independently and identically normally distributed with mean value of 0 and
variance of σt square.
• In the classical normal linear regression model it is assumed that
• that is, homoscedastic variance.
• But to take into account the ARCH effect, the following model is formed

• --------------------------

• That is, we assume that the error variance at time t is equal to some constant
plus a constant multiplied by the squared error term in the previous time
period.
• Of course, if λ1 is zero, the error variance is homoscedastic, in which case we
work in the framework of the classical normal linear regression model.
• It is assumed that the coefficients in this equation are positive because the
variance cannot be a negative number.
• Equation (15.5) is known as the ARCH (1) model, for it includes only one lagged
squared value of the error term.
• But this model can be easily extended to an ARCH (p) model, where we have p
lagged squared error terms, as follows:
• ---------------------------
• If there is an ARCH effect, it can be tested by the statistical significance of the
estimated λ coefficients
• If it significantly different from zero, we can conclude that there is an ARCH
effect.
• To test the ARCH effect in (15.6), we can use the F test to test the hypothesis
that

• Since the us are not directly observable, we first estimate Eq. (15.3) and
estimate u as
• and then estimate the following model:
• ------------ (15.10)
• That is, we regress the squared residuals at time t on its lagged values going up
to p previous period, the value of the p being determined empirically.
• Notice that in practice we replace which is replaced by its estimate
• As you can see, the AR part of the ARCH model is so-called because in Eq. (15.10)
we are regressing squared residuals on its lagged values going back to p periods.
• The CH part of ARCH is because variance in Eq. (15.10) is conditional on the
information available up to time (t – 1).
Estimation of the ARCH model: the
maximum likelihood approach
• The first part of the table gives the estimate of the mean equation and the
second half gives estimates of the coefficients of the variance equation.
• all the lagged variance coefficients are positive, as expected; the first three
coefficients are not individually statistically significant, but the last five are.
• It seems there is an ARCH effect in the dollar/euro exchange rate return. That is,
the error variances are autocorrelated.
The GARCH model
• Some of the drawbacks of the ARCH (p) model are as follows:
• first, it requires estimation of the coefficients of p autoregressive terms, which
can consume several degrees of freedom.
• Secondly, it is often difficult to interpret all the coefficients, especially if some of
them are negative.
• Thirdly, the OLS estimating procedure does not lend itself to estimate the mean
and variance functions simultaneously.
• Therefore, the literature suggests that an ARCH model higher than ARCH (3) is
better estimated by the GARCH model (Generalized Autoregressive Conditional
Heteroscedasticity) model
• In its simplest form, in the GARCH model we keep the mean equation (15.3) the
same, but modify the variance equation as follows:
• -------------(15.11)
• Notice that here the conditional variance at time t depends not only on the
lagged squared error term at time (t – 1) but also on the lagged variance term at
time (t – 1). This is known as the GARCH (1,1) model.
• ARCH (p) model is equivalent to GARCH (1, 1) as p increases.
• In the ARCH (p) given in Eq. (15.6) we have to estimate (p + 1) coefficients,
whereas in the GARCH (1,1) model given in Eq. (15.11) we have to estimate only
three coefficients.
• The GARCH (1,1) model can be generalized to the GARCH (p,q) model with p lagged
squared error terms and q lagged conditional variance terms, but in practice GARCH
(1,1) has proved useful to model returns on financial assets.
• Comparing the ARCH (8) with GARCH (1,1), we see how GARCH (1,1) in effect captures
the eight lagged squared error terms in Table 15.2.
• As you can see in the table below , in the variance equation both the lagged squared
error term and the lagged conditional variance term are individually highly significant.
• Since lagged conditional variance affects current conditional variance, there is clear
evidence that there is a pronounced ARCH effect.
• To sum up, there is clear evidence that the dollar/euro exchange rate returns exhibit
considerable time-varying and time-correlated volatility, whether we use the ARCH or
the GARCH model.