Deterministic and Stochastic Trends

Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Time Series Analysis

Models with Trends

Jude C. Hays

February 16, 2021

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Outline

1 Deterministic and Stochastic Trends

2 Spurious Regressions

3 Dickey-Fuller Tests

4 Univariate Decompositions

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Deterministic and Stochastic Trends

Time series processes can be decomposed into three parts:

the trend, the stationary component, and noise.
The trend component accounts for changes in the nature of
the time series over time.
Time series processes with trends are non-stationary. The
mean, variance, or both are a function of time.
We need to properly account for trends in dynamic processes
in order to test hypotheses with time series data.

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Deterministic Trends
A deterministic trend is one where realizations of the time
series process are a fixed function of time, such as a
high-order polynomial
yt = β0 + β1 t + β2 t 2 + β3 t 3
Clearly, in this case, E (yt ) depends on t.
If we add a stationary component to the trend, for example,
yt = β0 + β1 t + β2 t 2 + β3 t 3 + A(L)εt
The process is said to be trend-stationary. Long-run
forecasts will converge to the trend.
In the simplest case, we have yt = β0 + β1 t, a linear trend,
which can be expressed as
∆yt = β0 , or ∆yt = β0 + εt (with noise)
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Stochastic Trends
A stochastic trend is one where realizations of a random
process have permanent effects on the nature of a time series.
In the simplest case, we have a random walk process
yt = yt−1 + εt or ∆yt = εt
The solution to this first-order difference equation is
t
X
yt = y0 + εi
i=1

Hence, the effects of stochastic shocks do not decay over time.

This means that the variance of a random walk is time
dependent
var(yt ) = var(εt + εt−1 + ... + ε1 ) = tσ 2
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Mixed Trends
A random walk with drift
yt = a0 + yt−1 + εt or ∆yt = a0 + εt
has a trend that is partially deterministic and partially
stochastic.
Pt
The solution to this model is yt = y0 + a0 t + εi
i=1
Generalizations include the trend plus noise and trend plus
irregular models
t
P
yt = y0 + a0 t + εi + ηt
i=1
and
t
P
yt = y0 + a0 t + εi + A(L)ηt
i=1

respectively, where ηt is white noise.

Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Detrending and Differencing

In order to do time series analysis we need to distinguish the
trend and stationary components, and the appropriate method
depends on whether the trend is deterministic or stochastic.
If a trend is stochastic, we difference the data to isolate the
stationary component. The process is difference-stationary.
In the case of a random walk with drift, we have
E (∆yt ) = E (a0 + εt ) = a0
var(∆yt ) = E (∆yt − a0 )2 = E (εt )2 = σ 2
cov(∆yt , ∆yt−s ) = E (εt , εt−s ) = 0
If the trend is deterministic, to isolate the stationary
component, we detrend the data by regressing {yt } on a
high-order polynomial function of time.
The order of the polynomial can be determined by t-tests and
F -tests as well AIC and SBC measures of fit.
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Spurious Regressions

It should be clear that two times series with deterministic

trends will correlate spuriously.
If the true data generating process is

yt = fy (t) + εyt
zt = fz (t) + εzt

and we estimate the regression

yt = βzt + eyt ,

then eyt will contain fy (t), which will correlate with zt through
fz (t). The estimate of β will suffer from omitted variable bias.
Moreover, because the {eyt } are not independent, our
standard error estimates will be biased as well.
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Spurious Regressions
What if two times series have stochastic trends? Will they
also correlate spuriously?
The answer in this case is not as obvious. However, Granger
and Newbold (1974) showed that if the true data generating
process is
yt = yt−1 + εyt
zt = zt−1 + εzt
and we estimate the regression
yt = β0 + β1 zt + eyt
we will reject the null hypothesis more often than suggested
by our p-values. In their Monte Carlo simulations, t-tests
rejected the null hypothesis β = 0 about seventy-five
percent of the time.
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Dickey-Fuller Tests

Detecting purely deterministic trends is relatively easy. F and

t-tests will work.
Detecting purely stochastic and mixed trends is more
complicated.
We cannot rely on ACF plots. In finite samples, the ACF of
an integrated process will look like a stationary
near-integrated process.
Dickey and Fuller developed their tests around three
equations:
∆yt = γyt−1 + εt
∆yt = a0 + γyt−1 + εt
∆yt = a0 + γyt−1 + a2 t + εt

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Dickey-Fuller Tests

Dickey and Fuller developed their tests around three

equations:
∆yt = γyt−1 + εt
∆yt = a0 + γyt−1 + εt
∆yt = a0 + γyt−1 + a2 t + εt
Under the null hypothesis γ = 0, the first equation is a pure
random walk. The second equation adds a drift term, and the
third equation adds a drift and linear time trend.
Note that under the null hypothesis. The process is
non-stationary, the variance of yt is a function of time and
becomes infinitely large as t increases.
Thus, standard t and F -tests, which assume constant and
finite variance, are not appropriate.
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Dickey-Fuller Tests

A Dickey-Fuller test is conducted by estimating one of the

three regression, computing the relevant t and F -statistics
and comparing these against the empirical critical values
identified via simulation under the null hypothesis.

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Extensions: Augmented Dickey-Fuller Test

What if the time series process is a higher order autoregressive

process?

yt = a0 + a1 yt−1 + a2 yt−2 + ... + ap yt−p + εt

The augmented Dickey-Fuller test uses the regression

p
X
∆yt = a0 + γyt−1 + βi ∆yt−i+1 + εt
i=2

p p
 
P P
where γ = − 1 − ai and βi = − aj .
i=1 j=i

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Extensions: Perron’s Test

What if there’s a structural break in the time series?
Perron’s test allows us to distinguish a random walk with a
pulse from a trend-stationary process with a structural break
in the intercept.
H1 : yt = a0 + yt−1 + µ1 Dp + εt
A1 : yt = a0 + a2 t + µ2 DL + εt
The test proceeds as follows:
1 Detrend the data using the alternative model with residuals ŷt .
2 Estimate the regression ŷt = a1 ŷt−1 + εt .
3 Check for serial correlation and estimate
k
P
ŷt = a1 ŷt−1 + βi ∆ŷt−i + εt , if needed.
i=1
4 Calculate the t-statistic for the null hypothesis a1 = 1 and
compare against Perron’s table of critical values.
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Beveridge and Nelson Decomposition

If a time series has a stochastic trend plus either noise or an
irregular part, it may be interesting to decompose the series
into its permanent (trend) and temporary (stationary)
components.
One method for doing this is the Beveridge and Nelson
decomposition, which can be used for any ARIMA (p,1,q)
model.
Start by considering an ARIMA (0,1,2) process
yt = a0 + yt−1 + εt + β1 εt−1 + β2 εt−2
The solution for yt is
t
X
yt = a0 t + y0 + ei
i=1

where et = εt + β1 εt−1 + β2 εt−2 .

Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Beveridge and Nelson Decomposition

t+s
P
The solution for yt+s is a0 (t + s) + y0 + ei
i=1
Substitute for y0 using the solution for yt
 t
 t+s
P P
yt+s = a0 (t + s) + yt − a0 t − ei + ei
i=1 i=1
s
P
= a0 s + yt + et+i
i=1

Rewriting in terms of εt , we have

s
X s
X s
X
yt+s = a0 s + yt + εt+i + β1 εt−1+i + β2 εt−2+i
i=1 i=1 i=1

Through recursive substitution, it can be shown that

Et yt+s = a0 s + yt + (β1 + β2 )εt + β2 εt−1

Jude C. Hays Time Series Analysis

 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Beveridge and Nelson Decomposition

The time-t forecast of yt+s is the current level of the
stochastic trend plus the forecast of the deterministic trend:
Et yt+s = µt + a0 s. Solving for µt gives
µt + a0 s = yt + a0 s + (β1 + β2 )εt + β2 εt−1
or
µt = yt + (β1 + β2 )εt + β2 εt−1
Since the temporary component of yt is yt − µt − a0 t, we have
Temporary = yt − [yt − a0 t + (β1 + β2 )εt + β2 εt−1 ]
= −(β1 + β2 )εt − β2 εt−1
Hence, the difference between the observed {yt } and the
stochastic trend is perfectly negatively correlated with the
temporary part of {yt }. This is the identifying assumption
that allows us to perform the decomposition.
The decomposition proceeds in five steps.
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Beveridge and Nelson Decomposition

1 Difference yt and estimate a0 , β1 , and β2 as the parameters of
a stationary ARMA(0,2) model.
2 Use the ARMA model to make in-sample forecasts of each yt
and yt−1 .
3 Set the forecast errors equal to εt and εt−1 respectively.
ε0 = y0 − a0
ε1 = y1 − y0 − a0 − β1 ε0
ε2 = y2 − y1 − a0 − β1 ε1 − β2 ε0
..
.

4 Solve for the irregular component (β1 + β2 )εt + β2 εt−1

5 The trend is µt = (β1 + β2 ) εt + β2 εt−1 + yt and the transitory
component is yt − µt .
Jude C. Hays Time Series Analysis
 Deterministic and Stochastic Trends
Spurious Regressions
Dickey-Fuller Tests
Univariate Decompositions

Hodrick and Prescott Decomposition

The Hodrick-Prescott filter chooses µt to minimize the

following sum of squares
T T −1
1 X 2 λ X
(yt − µt ) + [(µt+1 − µt ) − (µt − µt−1 )]2
T T
t=1 t=2

Note that when λ is zero, the solution is µt = yt , and when

λ → ∞, the result is a linear time trend.
Hence, high values of λ smooth the time series.
There is also an unobserved components decomposition,
which we will cover when we get to state-space models.

Jude C. Hays Time Series Analysis