Characterizing Cycles

ECON3086: Topics in Econometrics

Dr Muhammad Shafiullah

School of Economics
University of Nottingham Malaysia

7 November 2021
Relevant book chapters

Chapters 2 & 7, “Elements of Forecasting”, 4th edition, Francis X.

Diebold, Cengage Learning.
Chapter 7, “Forecasting: Methods and Applications”, 3rd Edition,
Spyros Makridakis, Steven C. Wheelwright and Rob J. Hyndman,
John Wiley & Sons, Inc.
Chapter 1, 3 & 6, “Forecasting for Business and Economics”, 1st
Edition, Gloria González-Rivera, Pearson Education, Inc.

MS (UNM) Cycles (Part 1) 7 November 2021 2 / 32

List of Topics

1 Covariance Stationary Time Series.

2 White Noise.
3 The Lag Operator.
4 Estimation and Inference for the Mean, Autocorrelation, and Partial
Autocorrelation Functions.

MS (UNM) Cycles (Part 1) 7 November 2021 3 / 32

What is Cycle?

When you think of a ”cycle,” you probably think of the sort of rigid
up and down pattern depicted in Figure 7.1 (slide 5) & 1.4 (slide 6).
Such cycles can sometimes arise, but cyclical fluctuations in business,
finance, economics, and government are typically much less rigid.

MS (UNM) Cycles (Part 1) 7 November 2021 4 / 32

What is cycle? (contd.)

Source: Diebold (2007), Chapter 7, p.113.

MS (UNM) Cycles (Part 1) 7 November 2021 5 / 32

What is cycle? (contd.)

Source: González-Rivera (2013), Chapter 1, p.9.

MS (UNM) Cycles (Part 1) 7 November 2021 6 / 32

What is cycle? (contd.)

However, when we speak of cycles in time series forecasting, we have

in mind a much more general, all-encompassing notion of cyclicality:
any sort of dynamics not captured by trends or seasonals.
What economic variables follow a cyclical pattern, as well as trend
and seasonality?

MS (UNM) Cycles (Part 1) 7 November 2021 7 / 32

Covariance Stationary Time Series

If we want to forecast a series, at a minimum we’d like its mean and

its covariance structure (i.e., the covariances between current and
past values) to be stable over time, in which case we say that the
series is covariance stationary. (Covariance here is also referred to
as ”autocovariance” in Topic 1, Slide 19.)
Quantifying stability of the covariance structure is done using the
autocovariance function.The autocovariance at displacement τ is just
the covariance between yt and yt−τ . It will of course depend on τ ,
and it may also depend on t, so in general we write

γ(t, τ ) = cov (yt , yt−τ ) = E (yt − µ)(yt−τ − µ)

MS (UNM) Cycles (Part 1) 7 November 2021 8 / 32

Covariance Stationary Time Series (contd.)

Displacement here refers to the order of the observations in the time

series.
If the covariance structure is stable over time, as required by
covariance stationarity, then the autocovariances depend only on
displacement, τ , not on time, t, and we write

γ(t, τ ) = γτ

for all t.
This is related to the idea of ‘weak dependence’ in AE II.

MS (UNM) Cycles (Part 1) 7 November 2021 9 / 32

Covariance Stationary Time Series (contd.)
Although many series are not covariance stationary, it is frequently
possible to work with models that give special treatment to nonstationary
components such as trend and seasonality, so that the cyclical component
that’s left over is likely to be covariance stationary.

Source: González-Rivera (2013), Chapter 3, p.64.

MS (UNM) Cycles (Part 1) 7 November 2021 10 / 32

Covariance Stationary Time Series (contd.)

As correlations are easier to interpret than covariances, we often work

with the correlation, rather than the covariance, between yt and yt−τ .
That is, we work with the autocorrelation function, ρ(τ ), rather
than the autocovariance function, γ(τ ). The autocorrelation function
is obtained by dividing the autocovariance function by the variance.

γ(τ )
ρ(τ ) = , τ = 0, 1, 2, ...
γ(0)

Autocorrelation coefficients are basically how well your current values

are correlated with your past values.
E.g. when someone says I’m not the person I used to be—very low
autocorrelation coefficient.

MS (UNM) Cycles (Part 1) 7 November 2021 11 / 32

Covariance Stationary Time Series (contd.)

The formula for the autocorrelation is just the usual correlation

formula, specialized to the correlation between yt , and yt−τ . Note
that the variance of y , is γ(0), and by covariance stationarity, the
variance of y at any other time yt−τ is also γ(0). Thus,

cov (yt , yt−τ ) γ(τ ) γ(τ )

ρ(τ ) = p p =p p =
var (yt ) var (yt−τ ) γ(0) γ(0) γ(0)

Note that we always have ρ(0) = γ(0)γ(0) = 1, because any series is

perfectly correlated with itself. Thus, the autocorrelation at
displacement 0 isn’t of interest; rather, only the autocorrelations
beyond displacement 0 inform us about a series’ dynamic structure.

MS (UNM) Cycles (Part 1) 7 November 2021 12 / 32

Covariance Stationary Time Series (contd.)

Finally, the partial autocorrelation function, p(τ ), is sometimes useful.

p(τ ) is just the coefficient of yt−τ in a population linear regression of
yt on yt−1 , ..., yt−τ . We call such a regression an autoregression
(AR), because the variable is regressed on lagged values of itself.

yt = β0 + β1 yt−1 + β2 yt−2 + ... + βτ yt−τ + εt

(Recall Topic 2, Slide 25)

The autocorrelations and partial autocorrelations are related but differ
in an important way.

MS (UNM) Cycles (Part 1) 7 November 2021 13 / 32

Covariance Stationary Time Series (contd.)

The autocorrelations are just the ”simple” or ”regular” correlations

between yt and yt−τ . The partial autocorrelations, on the other hand,
measure the association between yt and yt−τ after controlling for the
effects of yt−1 , ..., yt−τ +1 ; that is, they measure the partial correlation
between y , and yt−τ .
In other words, it is necessary to run an autoregressive equation of yt
on yt−1 , ..., yt−τ +1 to obtain the partial autocorrelations whereas
autocorrelations can be obtained without running any regressions.
Remember from AEI that the coefficients of a multiple regression
”partial out” effects of the particular regressor on the regressand; the
horrid Ballantine Venn diagrams?

MS (UNM) Cycles (Part 1) 7 November 2021 14 / 32

Covariance Stationary Time Series (contd.)

We often graph the autocorrelations and partial autocorrelations as a

function of τ and examine their qualitative shape.

Source: Diebold (2007), Chapter 7, p.117.

MS (UNM) Cycles (Part 1) 7 November 2021 15 / 32

White Noise

Before we estimate time series forecasting models, we need to

understand their population properties, assuming that the postulated
model is true.
The simplest of all such time series processes is the fundamental
building block from which all others are constructed. In fact, it’s so
important that we introduce it now. We use y to denote the observed
series of interest. Suppose that

yt = εt

εt ∼ (0, σ 2 ),
where the ”shock,” εt , is uncorrelated over time. We say that εt , and
hence yt , is serially uncorrelated.

MS (UNM) Cycles (Part 1) 7 November 2021 16 / 32

White Noise (contd.)
Such a process, with zero mean, constant variance, and no serial
correlation, is called zero-mean white noise, or simply white noise.

εt ∼ WN(0, σ 2 ),

and hence
yt ∼ WN(0, σ 2 ),
Note that, although εt and hence yt are serially uncorrelated, they are
not necessarily serially independent, because they are not necessarily
normally distributed.
If in addition to being serially uncorrelated, y is serially independent,
then we say that y is independent white noise. We write

yt ∼iid (0, σ 2 )

y is independently and identically distributed with zero mean &

constant variance.
MS (UNM) Cycles (Part 1) 7 November 2021 17 / 32
White Noise (contd.)

The unconditional mean and variance of yt ∼iid (0, σ 2 ) are E (yt ) = 0 and
var (yt ) = σ 2 , respectively. The unconditional mean and variance are
constant. In fact, the unconditional mean and variance must be constant
for any covariance stationary process.

Source: Diebold (2007), Chapter 7, p.120.

MS (UNM) Cycles (Part 1) 7 November 2021 18 / 32

White Noise (contd.)
The autocovariance function for a white noise process is
(
σ2, τ = 0
γ(τ ) =
0, τ ≥1
the autocorrelation function for a white noise process is
(
1, τ = 0
ρ(τ ) =
0, τ ≥ 1

Source: Diebold (2007), Chapter 7, p.121.

MS (UNM) Cycles (Part 1) 7 November 2021 19 / 32

White Noise (contd.)

The partial autocorrelation function for a white noise process is

(
1, τ = 0
p(τ ) =
0, τ ≥ 1

It’s degenerate and exactly the same as the autocorrelation function!

Source: Diebold (2007), Chapter 7, p.121.

MS (UNM) Cycles (Part 1) 7 November 2021 20 / 32

The Lag Operator

The lag operator and related constructs are the natural language in
which forecasting models are expressed.
It ”operates” on a series by lagging it. Hence,

Lyt = yt−1

Similarly, a second order lag is,

L2 yt = L(L(yt )) = Lyt−1

and so on.
A well-known operator is the first-difference operator ∆,

∆yt = (1 − L)yt = yt − yt−1

MS (UNM) Cycles (Part 1) 7 November 2021 21 / 32

Wold’s representation theorem

The Wold decomposition is a theoretical result concerning only covariance

stationary processes. However, it is very powerful because regardless of
how the true process has been generated, the theorem claims that there is
always a linear approximation to the true process, which implies that at
least we will be able to construct a linear forecast. Wold’s representation
theorem points to the appropriate model—for what’s left after fitting the
trend and seasonal components.

MS (UNM) Cycles (Part 1) 7 November 2021 22 / 32

Wold’s representation theorem (contd.)

Theorem
Let yt be any zero-mean covariance-stationary process. Then we can write
it as
X∞
yt = Ψ(L)εt = ψi εt−i
i=0

εt ∼ WN(0, σ 2 ),
P∞ 2
where ψ0 = 1 and i=0 ψi < ∞.

ln short, the correct ”model” for any covariance stationary series is

some infinite distributed lag of white noise, called the Wold
representation.

MS (UNM) Cycles (Part 1) 7 November 2021 23 / 32

Wold’s representation theorem (contd.)

On average, the (squared) distance between a linear model and the

true model is negligible when we account for all the past information.
The εt ’s are often called innovations, because they correspond to the
1-step-ahead forecast errors that we’d make if we were to use a
particularly good forecast.
The εt ’s represent that part of the evolution of y that’s linearly
unpredictable on the basis of the past of y . Note also that the εt ’s,
although uncorrelated, are not necessarily independent.

MS (UNM) Cycles (Part 1) 7 November 2021 24 / 32

Wold’s representation theorem (contd.)

The Wold decomposition is very important because it guarantees that

there is always a linear model that can represent the dynamics of a
covariance stationary stochastic process.
There may be other more complicated representations, such as
nonlinear models, but we can always approximate the stochastic
process by a linear representation. This is the starting point in any
time series modeling strategy.

MS (UNM) Cycles (Part 1) 7 November 2021 25 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions
Portmanteau Tests

Rather than study the ρ̂(τ ) values one at a time, an alternative

approach is to consider a whole set of ρ̂(τ ) values, say the first 15 of
them (ρ̂(1) through ρ̂(15)) all at one time, and develop a test to see
whether the set is significantly different from a zero set.
Tests of this sort are known as portmanteau tests.
A common portmanteau test is the Box-Pierce test which is based
on the Box-Pierce Q-statistic:
m
X
QBP = T ρ̂2 (τ )
τ =1

where m is the maximum lag being considered and T is the number

of observations in the series. Usually m ≈ 20 is selected.

MS (UNM) Cycles (Part 1) 7 November 2021 26 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions (contd.)
Portmanteau Tests

The Box-Pierce test was designed by Box and Pierce (1970) for
testing the residuals from a forecast model.
If each ρ̂(τ ) is close to zero, QBP will be relatively small whereas if
some ρ̂(τ ) values are large (either positive or negative), the QBP
statistic will be relatively large.
QBP is approximately distributed as a χ2m random variable under the
null hypothesis that y is white noise.

MS (UNM) Cycles (Part 1) 7 November 2021 27 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions (contd.)
Portmanteau Tests

An alternative portmanteau test is the Ljung-Box test due to Ljung

and Box (1978).
m  
X 1
QLB = T (T + 2) ρ̂2 (τ )
T −τ
τ =1

They argued that the alternative statistic has a distribution closer to

the χ2 distribution than does the QBP statistic.
Under the null hypothesis that y is white noise, QLB is approximately
distributed as a χ2m random variable.

MS (UNM) Cycles (Part 1) 7 November 2021 28 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions (contd.)
Durbin-Watson Test

The Durbin-Watson statistic tests for correlation over time, called

serial correlation, in regression disturbances.
The Durbin-Watson test works within the context of the model:

yt = β0 + β1 xt + β2 zt + εt

εt = φεt−1 + vt
vt ∼iid N(0, σ 2 )
The regression disturbance is serially correlated when φ ̸= 0. The
hypothesis of interest is that φ = 0.
When φ = 0, the ideal conditions hold, but when φ ̸= 0, the
disturbance is serially correlated.

MS (UNM) Cycles (Part 1) 7 November 2021 29 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions (contd.)
Durbin-Watson Test

If φ > 0, the disturbance is positively serially correlated, and if φ < 0,

the disturbance is negatively serially correlated. Positive serial
correlation is typically the relevant alternative in the applications
that will concern us.
The formula for the Durbin-Watson (DW) statistic is
PT 2
i=2 (et − et−1 )
DW = PT 2
i=1 et

DW takes values in the interval [0,4], and if all is well, DW should be

around 2.

MS (UNM) Cycles (Part 1) 7 November 2021 30 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions (contd.)
Durbin-Watson Test

If DW is substantially less than 2, there is evidence of positive serial

correlation. As a rough rule of thumb, if DW is less than 1.5, there
may be cause for alarm, and we should consult the tables of the DW
statistic, available in many statistics and econometrics texts.
If DW is greater than 2, successive error terms are, on average, much
different in value from one another, i.e., negatively correlated.

MS (UNM) Cycles (Part 1) 7 November 2021 31 / 32

Estimation and Inference for the Mean, Autocorrelation,
and Partial Autocorrelation Functions (contd.)

The example in Chapter 7 of Diebold (2007) shows Canadian Employment

(pp.130-132). What do the ACF, PACF and Portmanteau test estimates
indicate? Is Canadian Employment White Noise? Why or why not?

MS (UNM) Cycles (Part 1) 7 November 2021 32 / 32