MSF 566 Time Series Analysis

Topic 04 1
 Stylized facts of economic time series

 ARCH process

Topic 04 2
 Many economic time series do not have a constant
mean and most exhibit phases of relative tranquility
followed by periods of high volatility

 Much of the current econometric research is

concerned with extending the Box-Jenkins
methodology to analyze these types of time-series
variables

Topic 04 3
 Examine the so-called stylized facts concerning the
properties of economic time-series data
◦ A stochastic variable with a constant variance is called
homoskedastic as opposed to heteroskedastic

◦ For series exhibiting volatility, the unconditional variance

may be constant even though the variance during some
periods is usually large

Topic 04 4
 Formalize simple models of variances exhibiting
heteroskedasticity
◦ Assets holders are interested in the volatility of returns
over the holding period, not over some historical period

◦ This forward-looking view of risk means that it is

important to be able to estimate and forecast the risk
associated with holding a particular asset
 We will develop the tools necessary to model and forecast
conditional heteroskedasticity

Topic 04 5
 Analyze a number of variants of the basic model for
conditional volatility
◦ The simplest model posits that the conditional
hetroskedasticity can be estimated as an autorregressive
process (ARCH)
 We will examines a number of important generalizations of the
basic ARCH model

Topic 04 6
 Most of the series contain a clear trend
◦ Real GDP and consumption (Figure 3.1) exhibit a upward
trend
◦ Real investment and government purchases (Figure 3.2)
are more volatile than either real GDP or consumption

 Shocks to a series can display a high degree of

persistence
◦ Neither of the interest rate series (Figure 3.3) has a clear
upward or downward trend
◦ Both show a high degree of persistence

Topic 04 7
Upward trend

Topic 04 8
More volatile than real GDP

Topic 04 9
No clear upward or downward trend
High degree of persistence

Topic 04 10
 The volatility of many series is not constant over time
◦ From Figure 3.4, we can see periods where the stock market
seems tranquil alongside periods with large increases and
decreases in the market
 Such series are called conditionally heteroskedastic if the
unconditional (or long-run) variance is constant, but there are
periods in which the variance is relatively high

 Some series seem to meander

◦ The real effective exchange rates (Figure 3.5) show no particular
tendency to increase or decrease
 The real values of the US and the Canadian dollar seem to go through
sustained periods of appreciation and then depreciation with no
tendency to revert to a long-run mean
 This type of “random walk” behavior is typical of nonstationary series

Topic 04 11
The volatility is not constant

Topic 04 12
 Some series share co-movements with other series
◦ (Figure 3.3) The federal funds rate and 10-year yield on US
government securities show no tendency to revert to a
long-run mean
◦ It is also clear that the two series never drift too far apart

◦ (Figure 3.6) it is not clear whether the various indexes of

real GDP levels share the same trend
 The movements in the three GDP series are such that all seem
to experience recessions and expansions simultaneously
 It is not clear whether the differences among the trend of
growth rate are statistically significant

Topic 04 13
No tendency to revert to long-run mean
-“random walk”

Topic 04 14
Not clear whether three series share Not clear whether the differences
the same trend, but they do tend to among the trend of growth rates
move simultaneously are statistically significant

Topic 04 15
 In conventional econometric models, the variance of
the disturbance term is assumed to be constant

 We have seen that many economic time series

exhibit period of unusually large volatility, followed
by period of relative tranquility
◦ The assumption of a constant variance (homoskedasticity)
in such circumstances is inappropriate

Topic 04 16
 Unconditional variance (i.e., the long-run forecast of
the variance) would be unimportant if you plan to
buy the asset at t and sell at t+1

 One approach to forecast the variance is to explicitly

introduce an independent variable that helps to
predict the volatility

Topic 04 17
  Consider the simplest case in which
yt 1   t 1 xt
◦ Where:

yt is the variable of interest

 t 1 is a white - noise disturbance term with variance  2
xt is an independen t variable that can be observed at period t

Topic 04 18
 If xt  xt 1  xt 2  ...  constant , the {yt} sequence is the
white-noise process with a constant variance

 When the realization of the {xt} sequence are not all

equal, the variance of yt+1 conditional on the
observable value of xt is var( yt 1 | xt )  xt2 2
◦ Here the conditional variance of yt+1 is dependent on the
realized value of xt

◦ If the magnitude of xt2 is large (small), the variance of yt+1

will be large (small) as well

Topic 04 19
 The methodology illustrated above forces {xt} to
affect both the mean and variance of yt

 Engle (1982) shows that it is possible to

simultaneously model the mean and the variance of
a series
◦ As a preliminary step to understand Engle’s method, note
that conditionally forecasts are vastly superior to
unconditional forecasts

Topic 04 20
 The unconditional forecast
has a greater variance than
the conditional forecast

 Suppose we are estimating the stationary ARMA

model yt  a0  a1 yt 1   t
◦ The conditional mean of yt+1 is Et yt 1  a0  a1 yt

◦ If we use this conditional mean to forecast yt+1, the forecast

error variance is Et [( yt 1  a0  a1 y t )2 ]  Et t21   2

◦ However, if unconditional forecast is used, the

unconditional forecast is always the long-run mean of the
{yt} sequence, that is a0/(1-a1)
 The unconditional forecast error variance is

E{[ yt 1  a0 /(1  a1 )]2 }  E[( t 1  a1 t  a12 t 1  a13 t 2  ...)2 ]
yt  a0 /(1  a1 )   a1 t i
i

i 0   2 /(1  a12 )
Topic 04 21
 We can estimate any tendency for sustained
movements in the variance using an ARMA model

 Let {ˆt } denote the estimated residuals from the

model yt  a0  a1 yt 1   t , so that the conditional
variance of yt+1 is
var( yt 1 | yt )  Et [( yt 1  a0  a1 yt )2 ]  Et ( t 1 )2
◦ To this point, we have set Et ( t 1 )2 equal to the constant σ2
◦ Now suppose that the conditional variance is not constant

Topic 04 22
 One simple strategy is to forecast the conditional
variance as an AR(q) process using squares of the
estimated residuals
ˆt2  0  1ˆt21  2ˆt22  ...   qˆt2q   t (*)
where  t is a white noise
◦ If all values of αi (i from 1 to q) are zero the estimated
variance is simply α0
◦ Otherwise, the conditional variance of yt evolves according
to the autoregressive process given by the above equation
(*)

Topic 04 23
 We can forecast the conditional variance at t+1 as
◦ For this reason, an equation like (*) is called autoregressive
conditional heteroskedastic (ARCH) model
Etˆt21  0  1ˆt2  2ˆt2  ...  qˆt2q

◦ There are many applications for ARCH models since the

residuals in (*) can come from an autoregressive, an ARMA
model, or a standard regression model

Topic 04 24
 ˆt2  0  1ˆt21  2ˆt22  ...   qˆt2q   t (*)
where  t is a white noise

 The linear specification in (*) is not the most

convenient
◦ The reason is that the model for {yt} and the conditional
variance are best estimated simultaneously using MLE

◦ Instead of the specification given by (*), it is more tractable

to specify νt as a multiplicative disturbance

Topic 04 25
 The simplest example proposed by Engle (1982) is
 t   t 0  1 t21 where  t is a white noise
 Such that 2  1 , νt and εt-1 are independent of each other; and α0
and α1 are constants such that α0>0 and 0< α1<1

Topic 04 26
 It is easy to show that the element of the {εt}
sequence have a mean of zero and are uncorrelated
◦ Take the unconditional expectation of εt
E t  E[ t ( 0  1 t21 )1 / 2 ]  E t E[( 0  1 t21 )1 / 2 ]  0
◦ Since Eνtνt-i=0, it also follows that E t t i  0 i  0

◦ Square εt and take unconditional expectation (E t2  E t21 )

E t2  E[ t2 (0  1 t21 )]  E t2 E (0  1 t21 )  0 /(1  1 )

 The unconditional mean and variance are unaffected by the

presence of the error process given by  t   t 0  1 t21

Topic 04 27
 It is easy to show the conditional mean of εt is equal
to zero
E ( t |  t 1,  t 2 ...)  Et 1 t Et 1 (0  1 t21 )1/ 2  0

 The influence of  t   t 0  1 t21 falls entirely on the

conditional variance
E[ t2 |  t 1,  t 2 ,...]  0  1 t21
◦ The conditional variance of εt is dependent on the realized
value of εt-12, if the realized value of  t 1 is large, the
2

conditional variance in t will be large as well

Topic 04 28
E[ |  t 1,  t 2 ,...]  0   
2
t
2
1 t 1

 The conditional variance follows a first-order

autoregressive process denote by ARCH(1)

 We have to restrict the coefficients

◦ The conditional variance is never negative
 It is necessary to assume both α0 and α1 are positive

◦ To ensure the stability condition of the process

 It is necessary to restrict α1 such that 0< α1<1

Topic 04 29
 In an ARCH model, the conditional and unconditional
expectation of the error terms are equal to zero

 The {εt} sequence is serially uncorrelated

E t t s  0 s  0
◦ The key point is that the errors are not independent since they
are related through their second moment (i.e., conditional
variance)

 The conditional variance itself is an autoregressive

process resulting in conditionally heteroskedastic errors
◦ Conditional heteroskedasticity in {εt} will result in {yt} being
hetroskedastic itself
The ARCH model is able to capture periods of
tranquility and volatility in the {yt} sequence
Topic 04 30
 In Figure 3.7, four panels depict two different ARCH
model
◦ Panel A resenting the {νt} sequence, shows 100 serially
uncorrelated and normally distributed random deviates
(white noise)

◦ Panel B uses the value in panel A to construct 100 values of

{εt}, setting α0=1 and α1=0.8
 The mean is zero
 The variance appears to experience an increase in the volatility
around t=50

Topic 04 31
 If we set a1=0, then for the model yt  a1 yt 1   t
◦ Panel B also depicts the time path of {yt} (i.e., yt   t )

 Panel C and D show the behavior of the {yt}

sequence for the case a1=0.2 and a1=0.9

 The essential point to note is that the ARCH error

structure and the autocorrelation parameters of the
{yt} process interact with each other
◦ Comparing Panel C and D illustrates that the volatility of
{yt} is increasing in a1 and α1

Topic 04 32
Topic 04 33
  The conditional mean and variance are given by
Et 1 yt  a0  a1 yt 1
var( yt | yt 1 , yt 2 ,...)  Et 1 ( yt  a0  a1 yt 1 )2  Et 1 t2  0  1 t21
 The unconditional mean and variance of yt can be
obtained by solving the differencing equation for yt
and then take expectations

Topic 04 34
 If the process began sufficiently far the past (the arbitrary
constant A can safely be ignored),the solution for yt is

a0 a
yt    a1i t i  Ey t  0
1  a1 i 0 1  a1

0 1
var( yt )   a12i var( t i )  ( )( )
i 0 1  1 1  a1
2

 Note that E t2  E t21  ...  0 /(1  1 )

 Clearly, the variance of {yt} sequence is increasing in α1 and in

the absolute value of a1
◦ Key point: an ARCH error process can be used to model periods of
volatility in the univariate framework

Topic 04 35
 Engle (1982)
◦ The entire class of higher-order ARCH(q) process
q
 t   t 0  i t2i
i 1

◦ All shocks from εt-1 to εt-q have a direct effect on εt, so that
the conditional variance acts like an autoregressive process
of order q

Topic 04 36