Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Non-stationary Time Series
Helle Bunzel
ISU
April 8, 2009
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Introduction to Non-Stationary Series I
What we have seen so far are series which can be written as
yt = µ+∞
∑j=0
ψj εtj = µ+ ψ (L) εt
where:
ψ0 = 1fεtg is a white noise process with mean 0 and variance σ2.The coe¢ cients are absolutely summable:
∞
∑j=0
ψj < ∞
The roots of ψ (L) = 0 are outside of the unit circle.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Introduction to Non-Stationary Series II
A couple of important feature of such series are:
E (yt ) = µ
and
lims!∞
yt+s jt = µ
This is not always desirable. Consider, for example US GDP:
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Introduction to Non-Stationary Series III
Clearly there is an upwards trend, which must be accounted for in anymeaningful forecast.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Non-Stationarity IThere are are two main approaches to modelling trends like these.The rst approach involves including a deterministic time trend:
yt = µ+ δt + ψ (L) εt
or in the case of the US GDP maybe
yt = µ+ δt + γt2 + ψ (L) εt
The other approach is to model the data as a unit root process:
(1 L) yt = δ+ ψ (L) εt
or
∆yt = δ+ ψ (L) εt
where it is assumed that ψ (1) 6= 0.This is also called a stochastic trend.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Non-Stationarity IIWhy did we need ψ (1) 6= 0?
Suppose the original fytg was a stationary series, say
yt = µ+ χ (L) εt
Then, when we di¤erence, we get:
∆yt = (1 L) χ (L) εt ψ (L) εt
where
ψ (L) = (1 L) χ (L)
In this case, we get:
ψ (1) = (1 1) χ (1) = 0
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Non-Stationarity III
Thus, ruling out ψ (1) = 0, essentially rules out that the original serieswas stationary. I.e. we had to di¤erence in order to obtain a stationaryseries.
Examples of simple unit root processes is the random walk withoutdrift:
yt = yt1 + εt
or the random walk with drift:
yt = δ+ yt1 + εt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Non-Stationarity IV
An alternative representation:
yt = α+ δt + ut
where ut is a mean 0 ARMA(p,q) process, such that1 φ1L φ2L
2 ... φpLput =
1+ θ1L+ θ2L2 + ...+ θqLq
εt
where the MA operator1+ θ1L+ θ2L2 + ...+ θqLq
is invertible.
Now, write the AR part as1 φ1L φ2L
2 ... φpLp= (1 λ1L) (1 λ2L) ... (1 λpL)
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Non-Stationarity V
If jλi j < 1 8i , then we can write
ut =
1+ θ1L+ θ2L2 + ...+ θqLq
(1 λ1L) (1 λ2L) ... (1 λpL)
εt ψ (L) εt
and
yt = α+ δt + ut
is a trend-stationary process.
Now suppose instead that λ1 = 1 and jλi j < 1 for i 2 f2, 3, ..., pg .(The series has a unit root)
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Non-Stationarity VIThen we can write
(1 L) ut =1+ θ1L+ θ2L2 + ...+ θqLq
(1 λ2L) ... (1 λpL)
εt ψ (L) εt
so
∆ut = ψ (L) εt
where ψ (L) satises the usual properties of stationary sequences.Now consider the original series:
yt = α+ δt + ut
Di¤erence the series:∆yt = yt yt1 = (α+ δt + ut ) (α+ δ (t 1) + ut1)
= δ+ ut ut1 = δ+ ∆ut = δ+ ψ (L) εtIn this situation, yt is a unit root process!
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some denitions I
When yt contains a unit root it is also said to be integrated of order1, or y I (1) .If yt contains two unit roots, such that
(1 L) (1 L) ut =
1+ θ1L+ θ2L2 + ...+ θqLq
(1 λ3L) ... (1 λpL)
εt
ψ (L) εt ,(1 L)2 ut = ψ (L) εt
then the data must be di¤erenced twice to become stationary:
(1 L) yt = ∆yt = δ+ ∆ut = δ+ (1 L) ut(1 L)2 yt = (1 L)∆yt = (1 L) [δ+ (1 L) ut ]
= (1 L)2 ut = ψ (L) εt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some denitions II
In this case we say that yt is integrated of order 2 or y I (2) .In general these process are called autoregressive integrated movingaverage processes or ARIMA (p, d , q) . Here p is the number of ARroots (other than the unit roots), d is the number of unit roots and qis the number of MA roots.
If you di¤erence an ARIMA (p, d , q) process d times, you produce astationary ARMA (p, q) process.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Economic Series I
Looking at GDP it seems that the best model might be exponentialgrowth, such that
yt = eδt
and
∂yt∂t= δeδt = δyt
This is called "proportional growth" and is common in economics.
In this case we simply take logs:
log (yt ) = δt
and then we can apply the determininstic trend model.Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Modelling Economic Series II
Similarly it is common to take logs before applying the unit rootmodel. This amounts to constant percentage of growth. For example:
Ination follows a stationary process.Population growth follows a stationary process.
To see this, consider:
∆ log (yt ) = logytyt1
= log
1+
yt yt1yt1
' yt yt1
yt1
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecasts I
First consider the trend-stationary series:
yt = µ+ δt + ψ (L) εt
Here the forcast is
yt+s jt = µ+ δ (t + s) + ψs εt + ψs+1εt1 + ....
Note that
Eyt+s jt µ+ δ (t + s)
2= E
ψs εt + ψs+1εt1 + ....
2 !s!∞
0
because the ψj are absolutely summable.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecasts IINow consider forecasting a unit root series
(1 L) yt = δ+ ψ (L) εt
Since ∆yt is a stationary process, we can apply the usual formulas to∆yt instead of yt .That implies that
∆yt+s jt = δ+ ψs εt + ψs+1εt1 + ....
Now suppose we want to forecast the level of yt+s . To do this, notethat
yt+s = yt+s yt+s1 + yt+s1 yt+s2 + ...+ yt+1 yt + yt= ∆yt+s + ∆yt+s1 + ∆yt+1 + yt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecasts III
Then the forecast is
yt+s jt = ∆yt+s jt + ∆yt+s1jt + ∆yt+1jt + yt=
δ+ ψs εt + ψs+1εt1 + ....
+
δ+ ψs1εt + ψs εt1 + ....
+
δ+ ψs2εt + ψs1εt1 + ....
+...
+ fδ+ ψ1εt + ψ2εt1 + ....g+yt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecasts IV
This implies that
yt+s jt = δs + yt +ψs + ψs1 + ...+ ψ1
εt
+ψs+1 + ψs1 + ...+ ψ2
εt1
+ψs+2 + ψs1 + ...+ ψ3
εt2 + ......
Note that the deterministic part is similar for the two types ofprocesses.
A few examples of unit root processes and their forcasts:
The random walk with drift:
yt = δ+ yt1 + εt
Here ψj = 0 for all j .
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecasts VThen
yt+s jt = δs + yt
An ARIMA (0, 1, 1) such that
yt = δ+ yt1 + εt + θεt1
and
ψ1 = θ
This implies that
yt+s jt = δs + yt + θεt
For both types of processes the forcast converges to a line:
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecasts VIFor the deterministic trend model:
yt+s jt = µ+ δ (t + s) + ψs εt + ψs+1εt1 + ....
!s!∞
µ+ δ (t + s) = (µ+ δt) + δs
For the unit root model:yt+s jt = δs + yt +
ψs + ψs1 + ...+ ψ1
εt
+ψs+1 + ψs1 + ...+ ψ2
εt1
+ψs+2 + ψs1 + ...+ ψ3
εt2 + ......
!s!∞
δs + yt
The slopes of these lines are identical, but the intercepts are not.
For the deterministic model the intercept is xed.For the unit root model the intercept varies with each new observation.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecast Errors I
First consider the trend-stationary process:
yt+s jt = µ+ δ (t + s) + ψs εt + ψs+1εt1 + ....
and
yt+s = µ+ δ (t + s) + εt+s + ψ1εt+s1 + ....
so
yt+s yt+s jt = fεt+s + ψ1εt+s1 + ....g
ψs εt + ψs+1εt1 + ...
= εt+s + ψ1εt+s1 + ....+ ψs1εt+1
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecast Errors II
The MSE is:
Ehyt+s yt+s jt
2i=1+ ψ21 + ψ22 + ...+ ψ2s1
σ2
and
lims!∞
Ehyt+s yt+s jt
2i= σ2
1+
∞
∑i=1
ψ2i
!
This is the same as the stationary MA processes.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecast Errors III
For the unit root process we get:
yt+s jt = δs + yt
+
ψs εt + ψs+1εt1 + ....
+
ψs1εt + ψs εt1 + ....
+
ψs2εt + ψs1εt1 + ....
+...
+ fψ1εt + ψ2εt1 + ....g
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecast Errors IV
and
yt+s = ∆yt+s + ∆yt+s1 + ∆yt+1 + yt= δs + yt
+ fεt+s + ψ1εt+s1 + ....g+ fεt+s1 + ψ1εt+s2 + ....g+...
+ fεt+1 + ψ1εt + ....g
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecast Errors VWe get
yt+s yt+s jt= fεt+s + ψ1εt+s1 + ....g
ψs εt + ψs+1εt1 + ....
+ fεt+s1 + ψ1εt+s2 + ....g
ψs1εt + ψs εt1 + ....
+...
+ fεt+1 + ψ1εt + ....g fψ1εt + ψ2εt1 + ....g=
εt+s + ψ1εt+s1 + ...+ ψs+1εt+1
+ fεt+s1 + ψ1εt+s2 + ....+ ψs εt+1g+...
+εt+1
= εt+s + (1+ ψ1) εt+s1 + (1+ ψ1 + ψ2) εt+s2 + ....
+1+ ψ1 + ψ2 + ...+ ψs1
εt+1
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Forecast Errors VI
This implies the MSE:
Ehyt+s yt+s jt
2i= σ2
h1+ (1+ ψ1)
2 + ...+1+ ψ1 + ψ2 + ...+ ψs1
2iThis one does not converge to a xed number. It can be shown that
Ehyt+s yt+s jt
2i 1+ ψ1 + ψ2 + ...+ ψs12
σ2s !s!∞
C
There is no upper bound on MSE!
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Dynamic Multipliers I
For the trend-stationary process:
yt+s = µ+ δ (t + s) + εt+s + ψ1εt+s1 + ....
and therefore
∂yt+s∂εt
= ψs
and
lims!∞
∂yt+s∂εt
= 0
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Dynamic Multipliers IIFor the unit root process:
yt+s = δs + yt+ fεt+s + ψ1εt+s1 + ....g+ fεt+s1 + ψ1εt+s2 + ....g+...
so
∂yt+s∂εt
= 1+ ψs + ψs1 + ...+ ψ1 = 1+s
∑i=1
ψi
and
lims!∞
∂yt+s∂εt
= 1+∞
∑i=1
ψi = ψ (1)
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Dynamic Multipliers III
A numerical example of this multiplier:
An ARIMA(4,1,0) was estimated for yt = 100 log (GDPt ) with theestimators:
∆yt = 0.555+ 0.312∆yt1+ 0.122∆yt2 0.116∆yt3 0.081∆yt4+ εt
Rewriting this as and MA process:1 0.312L 0.122L2 + 0.116L3 + 0.081L4
∆yt = 0.555+ εt
Dene φ (L) = 1 0.312L 0.122L2 + 0.116L3 + 0.081L4Check the stationarity of the estimated process: Turns out φ (L) hascomplex roots with norms 1.5 and 2.3.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Comparison of Dynamic Multipliers IV
The long-run e¤ect of a shock in εt is ψ (1) . Note that
ψ (L) =1
φ (L)
and therefore
lims!∞
∂yt+s∂εt
= ψ (1) =1
φ (1)=
1(1 0.312 0.122+ 0.116+ 0.081) = 1.31
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Transformations to Achieve Stationarity I
Note that if we have a trend-stationary process, then
yt δt = µ+ ψ (L) εt
is a stationary process.
Now, if the data is generated by a random walk with a drift, then
yt = δ+ yt1 + εt
yt δt = y0 + εt + εt1 + ...+ ε1
The mean is stationary, but the variance is σ2t.
Similarly if a trend-stationary process is di¤erenced, the resultingprocess is stationary, but a unit root is introduced in the MA processmaking it non-invertible.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Meaning of Unit Root tests I
Because of the inherent di¤erences between trend-stationary and unitroot processes, it is of interest whether an economic process is one orthe other.
Unfortunately, it is not possible to distinguish between the two withnite samples.
The two processes:
yt = yt1 + εt
and
yt = φyt1 + εt , jφj < 1
cannot be distinguished in nite samples, especially for φ very close to1.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Meaning of Unit Root tests IIA primary cause is that the processes only di¤er in their implicationsover innite horizons.
Consider the forecasts:
Unit root:
ys+t jt = yt
and
Eyt+s ys+t jt
2= sσ2
Trend-stationary:
ys+t jt = φsyt
Eyt+s ys+t jt
2=1+ φ2 + φ4 + ....+ φ2(s1)
σ2
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Meaning of Unit Root tests III
Clearly for φ su¢ ciently close to 1, the predictions will be virtuallyidentical.
What are the questions we can answer?
If we have enough data, we can nd out if a shock today has asignicant e¤ect 3 years from now.
While we wont know for sure which process generated the data, thismay be all we really need to know.
If we are willing to restrict ourselves to a specic AR process, e.g.AR(1):
yt = φyt1 + εt
then we can test H0 : φ = 1.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Meaning of Unit Root tests IV
It is probably the case that if φ0 = 0.998 we will not reject H0. Thatmay be OK, as the unit root process is a good approximation to thedata and will provide reasonable forecasts etc in nite time.
Note that this provides yet another argument for parsimonious models.
Witin a model, deciding whether to impose the unit root restriction isimportant for the usual reasons. E¢ ciency vs consistency.
Also the asymptotic theory used to test hypotheses di¤ers dependingon whether or not there is a unit root.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Asymptotic Theory for Dependent processes.
Almost all the asymptotic theory (LLN & CLT) you have seen hasfocussed on iid sequences.
At most you have seen heteroscedastic series.
Clearly this will not provide asymptotic theory for the kind ofprocesses that we have been working with.
This will be a brief introduction, not a thorough investigation.
The material can be found in Hamilton Chapter 7.2.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
LLN for stationary processes I
Assume that we have a covariance stationary process fYtg∞t=1 with
absolutely summable autocovariances.
Let
µ E (Yt )
γj = E [(Yt µ) (Ytj µ)]
Consider the sample mean:
YT =1T
T
∑t=1Yt
and note that
E (YT ) =1T
T
∑t=1
µ = µ
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
LLN for stationary processes II
If we can show that V (YT )! 0 as T ! ∞, then YT is consistentfor the mean. Note that weve been using this result.
Find V (YT ) :
V (YT )
= E
24 1T
T
∑t=1Yt µ
!235 = 1T 2E
24 T
∑t=1(Yt µ)
!235=
1T 2E
" T
∑t=1(Yt µ)
! T
∑s=1(Ys µ)
!#
=1T 2E
"T
∑t=1
T
∑s=1(Yt µ) (Ys µ)
#
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
LLN for stationary processes III
=1T 2
T
∑t=1
T
∑s=1
E [(Yt µ) (Ys µ)] =1T 2
T
∑t=1
T
∑s=1
γts
=1T 2fTγ0 + 2 (T 1) γ1 + 2 (T 2) γ2 + ...
+2 (T j) γj + ...+ 2 (T (T 1)) γT1
=1T 2
T1∑i=0
(T i) γi 1T 2
T1∑i=0
Tγi 1T
T1∑i=0
jγi j ! 0
We have now demonstrated ergodicity for the mean of a covariancestationary series with absolutely summable autocovariances.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some Denitions and Theorems I
Let fYtg∞t=1 be a sequence of random scalars with E (Yt ) = 0 for all
t.
Let Ωt be all the available information at time t.
Denition
If E (Yt jΩt1) = 0 for t = 2, 3, .... then fYtg is a martingale di¤erencesequence with respect to fΩtg .
Note: This is a stronger condition than fYtg being uncorrelated, butweaker than independence.
Example 1: If εt iid N0, σ2
and Yt = εt εt1, then
E (YtYt1) = Eεt ε
2t1εt2
= 0
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some Denitions and Theorems IINow consider
E (Yt jYt1,Yt2, ...,Y1)
If ε0 is a known constant, then this amounts to
E (Yt jεt1, εt2, ..., ε1) = E (εt εt1jεt1, εt2, ..., ε1)= E (εt ) εt1 = 0
If ε0 N0, σ2
and unknown, then Yt1,Yt2, ...,Y1 does not
provide information on εt1 and
E (Yt jYt1,Yt2, ...,Y1) = E (εt εt1jYt1,Yt2, ...,Y1)= E (εt εt1) = E (εt )E (εt1) = 0
So this sequence is uncorrelated, a martingale di¤erence sequence,but not serially independent.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some Denitions and Theorems III
Example 2: If εt iid N0, σ2
and Yt = ε2t εt1, then the series is
uncorrelated because:
E (YtYt1) = Eε2t ε
3t1εt2
= 0
Now consider
E (Yt jεt1, εt2, ..., ε1) = Eε2t εt1jεt1, εt2, ..., ε1
= E
ε2t
εt1 = σ2εt1
This, in general, is not 0, and hence Yt is not a martingale di¤erenceprocess.
Now for a more general denition:
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some Denitions and Theorems IV
Denition
If we can nd sequences of non-negative deterministic constants fctg∞t=1
and fξmg∞m=1 such that limm!∞ξm = 0 and
E jE (Yt jΩtm)j ctξm , 8t,m
then fYtg is said to follow an L1mixingale with respect to fΩtg .
Note that E (Yt jΩtm) is the m period ahead forecast, and that itmust converge to 0 as m! ∞.Hamilton provides a couple of examples.
With a few additional conditions, LLN for rst and second momentscan be proven for L1mixingales.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some Denitions and Theorems V
Furthermore, we have the following two CLTs:
Theorem
Let fYtg∞t=1 be a scalar martingale di¤erence sequence with
YT = 1T ∑T
t=1 Yt . Suppose that (a) EY 2t= σ2t > 0 with
1T ∑T
t=1 σ2t ! σ2 > 0, (b) E jYt jr < ∞ for some r > 2 and all t, and (c)1T ∑T
t=1 Y2t
P! σ2. ThenpTYT ) N
0, σ2
.
Note that Hamilton also provides denitions and theorems for vectors.
Clearly even when we have a scalar martingale di¤erence sequencethere is some work involved in showin that the conditions of the CLThold. The following proposition is useful for that purpose:
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Some Denitions and Theorems VI
Proposition Let Xt be a strictly stationary process withEX 4t= µ4 < ∞. Let Yt = ∑∞
j=0 hjXtj , where∑∞j=0 jhj j < ∞. Then Yt is a strictly stationary process with
E jYsYtYuYv j < ∞ for all s, t, u, v .
Next an example using many of these concepts.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
CLT Example I
Let fYtg∞t=1 be an MA (∞) process, such that Yt = ∑∞
j=0 ψj εtj ,
where ∑∞j=0
ψj < ∞.
Let εt be iid with E (εt ) = 0, Eε2t= σ2 and E
ε4t< ∞.
Now consider
Xt = εtYtk , k > 0
We will assume (c), namely that
1T
T
∑t=1X 2t
P! EX 2t
If you are interested Hamilton shows this, but it is non-trivial.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
CLT Example II
First verify that Xt is a martingale di¤erence sequence.
E (Xt jΩt1) = E (Xt jXt1,Xt2, ...)
Note that εt and Ytk are independent and that Xt1,Xt2, ...contains no information about εt , so
E (Xt jXt1,Xt2, ...) = E (εtYtk jXt1,Xt2, ...)= E (εt )E (Ytk jXt1,Xt2, ...)= 0
Hence Xt is a matingale di¤erence process.
We now need to verify conditions (a) and (b) of the CLT.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
CLT Example III
To verify condition (a) note that
EX 2t= E
ε2tY
2tk= E
ε2tEY 2tk
= σ2E
Y 2tk
We can nd
EY 2tk
= E
24 ∞
∑j=0
ψj εtj
!235 = σ2∞
∑j=0
ψ2j
Note that EY 2t= E
Y 2sfor all t, s.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
CLT Example IV
Since we have assume that ∑∞j=0
ψj < ∞, we also have
σ2 ∑∞j=0 ψ2j < ∞ and therefore
EX 2t= σ2E
Y 2tk
= σ4
∞
∑j=0
ψ2j < ∞
We now need to show that
limT!∞
1T
T
∑t=1EX 2t= lim
T!∞
1T
T
∑t=1
σ2EY 2tk
= σ2
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
CLT Example VThis follows easily from the fact that E
Y 2t= E
Y 2sfor all t, s :
limT!∞
1T
T
∑t=1EX 2t= lim
T!∞
1T
T
∑t=1
σ2EY 2tk
= lim
T!∞σ2E
Y 2t= σ4
∞
∑j=0
ψ2j = EX 2t
We can now apply the theorem, so we know that
pT1T
T
∑t=1Xt ) N
0,E
X 2t
or
1pT
T
∑t=1
εtYtk ) N0, σ2E
Y 2t
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
CLT Example VI
Now we turn to the properties of series with deterministic andstochastic trends.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model I
Consider the simple time trend model:
yt = α+ δt + εt ,
where εt is a white noise process.
Note: If εt is normal, all the usual nite sample results hold.
Asymptotics are inuenced by the non-stationarity, but nite sampletheory is not.
We will not assume normality and instead develop the asymptotictheory for this case.
Consider estimation with OLS and nd β.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model II
Let xt =1t
and β =
αδ
. Then
yt = x 0tβ+ εt
and
β =
T
∑t=1xtx 0t
!1 T
∑t=1xtyt
=
T
∑t=1xtx 0t
!1 T
∑t=1xtx 0tβ+ εt
= β+
T
∑t=1xtx 0t
!1 T
∑t=1xt εt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model III
We get:
β β =
T
∑t=1
1 tt t2
!1 T
∑t=1
εttεt
=
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
1 ∑Tt=1 εt
∑Tt=1 tεt
First consider ∑T
t=1 xtx0t :
T ∑Tt=1 t
∑Tt=1 t ∑T
t=1 t2
=
"T T (T+1)
2T (T+1)
2T (T+1)(2T+1)
6
#
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model IV
Note that
T1T ! 1
T2T (T + 1)
2! 1
2
T3T (T + 1) (2T + 1)
6! 1
3In fact, it turns out that
T(ν+1)T
∑t=1tυ ! 1
ν+ 1
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model V
Therefore, if we want to pre-multiply by some order of T such thatthe whole matrix converges, we need to multiply by T3. In that case:
T3
T ∑Tt=1 t
∑Tt=1 t ∑T
t=1 t2
=
"T3T T3 T (T+1)2
T3 T (T+1)2 T3 T (T+1)(2T+1)6
#!0 00 1
3
This matrix is not invertible. A trick that is used instead is to pre-and post multiply by Υ1T , where
ΥT =
"T
12 00 T
32
#
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model VI
Then we get:"T
12 0
0 T32
# T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
"T
12 0
0 T32
#
=
"T
12 0
0 T32
# "T
12T T
32 ∑T
t=1 tT
12 ∑T
t=1 t T32 ∑T
t=1 t2
#
=
"T
12T
12T T
12T
32 ∑T
t=1 tT
32T
12 ∑T
t=1 t T32T
32 ∑T
t=1 t2
#
=
T1T T2 ∑T
t=1 tT2 ∑T
t=1 t T3 ∑Tt=1 t
2
!1 1
212
13
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model VII
Now return to the estimator:
β β =
α α
δ δ
=
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
1 ∑Tt=1 εt
∑Tt=1 tεt
It turns out that we can pre-multiply
β β
by ΥT in exactly the
same way we pre-multiplied
β βbypT in the stationary model.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model VIII
If we do this, we get:
ΥT
β β
=
"T
12 00 T
32
# α α
δ δ
=
" pT (α α)
T32δ δ
#
= ΥT
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
1 ∑Tt=1 εt
∑Tt=1 tεt
=
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1 ∑Tt=1 εt
∑Tt=1 tεt
=
ΥTΥ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1 ∑Tt=1 εt
∑Tt=1 tεt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model IX
=
ΥTΥ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1 ∑Tt=1 εt
∑Tt=1 tεt
=
Υ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1Υ1T
∑Tt=1 εt
∑Tt=1 tεt
=
Υ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1 "T
12 ∑T
t=1 εtT
32 ∑T
t=1 tεt
#We know that
Υ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1!1 1
212
13
1=
4 66 12
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XIf we assume that E (εt ) = 0, V (εt ) = σ2 and E
ε2t< ∞, we can
apply a standard CLT:
T12
T
∑t=1
εt ) N0, σ2
The last bit we need is the limit of
T32
T
∑t=1tεt .
For this we will use the CLT for martingale di¤erence sequences. Firstnote that
T32
T
∑t=1tεt =
1pT
T
∑t=1
tT
εt
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XINow let xt = t
T εt and note that E (xt jΩt1) = 0.
Recall the theorem:
Theorem
Let fYtg∞t=1 be a scalar martingale di¤erence sequence with
YT = 1T ∑T
t=1 Yt . Suppose that (a) EY 2t= σ2t > 0 with
1T ∑T
t=1 σ2t ! σ2 > 0, (b) E jYt jr < ∞ for some r > 2 and all t, and (c)1T ∑T
t=1 Y2t
P! σ2. ThenpTYT ) N
0, σ2
.
We know that Ex2t= E
tT
2ε2t
= tT
2σ2 and that
1T
T
∑t=1
tT
2σ2 = σ2
1T 3
T
∑t=1t2 ! σ2
3
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XII
Also since we assumed that εt has nite fourth moments,
Ex4t= E
tT
4ε4t
!=
tT
4Eε4t< ∞
which satises (b).
Finally we need to show that
1T
T
∑t=1x2t
P! σ2
3
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XIII
Now, we know that
E
(1T
T
∑t=1x2t
)! σ2
3
so the proof is done if we can verify that
V
(1T
T
∑t=1x2t
)! 0
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XIV
Using the usual formula
V
(1T
T
∑t=1x2t
)
= E
24 1T
T
∑t=1x2t
1T
T
∑t=1Ex2t!235
= E
24 1T
T
∑t=1
tT
2ε2t
1T
T
∑t=1
tT
2σ2
!235= E
24 1T
T
∑t=1
tT
2 ε2t σ2
!235Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XV
V
(1T
T
∑t=1x2t
)
= E
24 1T
T
∑t=1
tT
2 ε2t σ2
!235=
1T 2
T
∑t=1
tT
4Eh
ε2t σ22i
=Eh
ε2t σ22i
T1T 5
T
∑t=1t4 ! 0 1
5= 0
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XVIThis veries that
1T
T
∑t=1x2t
P! σ2
3
and hence we can use the CLT, so
pTxT ) N
0,
σ2
3
or
1pT
T
∑t=1
tT
εt = T
32
T
∑t=1tεt ) N
0,
σ2
3
Now a quick re-cap:
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XVIIThe model is the simple time trend model:
yt = α+ δt + εt ,
We are considering the OLS estimator of the parameters.
So far we have shown that" pT (α α)
T32δ δ
#
=
Υ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1 "T
12 ∑T
t=1 εtT
32 ∑T
t=1 tεt
#where
ΥT =
"T
12 00 T
32
#Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XVIII
and Υ1T
T ∑T
t=1 t∑Tt=1 t ∑T
t=1 t2
Υ1T
1!1 1
212
13
1We also know the limiting distributions of T
12 ∑T
t=1 εt andT
32 ∑T
t=1 tεt . Some additional work is required to show their jointlimiting distribution (which we need).
It is fairly easy to show that this is Gaussian and that the covariance
matrix is1 1
212
13
σ2.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
The Simple Trend Model XIX
From this, we get that" pT (α α)
T32δ δ
# )1 1
212
13
1N0,1 1
212
13
σ2
= N
0,1 1
212
13
1σ2
!
= N0,4 66 12
σ2
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model I
Now that we have the distribution of the parameter estimates we canconsider testing hypotheses about the parameters.
First note thatpT (α α)) N
0, 4σ2
, so as long as the standard
estimate of the variance works, the standard t test for hypotheseson α should work. This turns out to be the case.
Now consider the standard t test for hypotheses on δ :
tT =
δ δ
hs2Tn(X 0TXT )
1o22
i 12
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model II
where xt =1t
,
XT =
2664x 01x 02:x 0T
3775and fgij signies the ij 0th entry of the matrix fg .
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model III
Re-writing the statistic:
tT =
δ δ
hs2Tn(X 0TXT )
1o22
i 12
=
δ δ
s2T0 1
(X 0TXT )
101
12
=T
32δ δ
s2Th0 T
32
i(X 0TXT )
10T
32
12
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model IV
=T
32δ δ
s2T0 1
ΥT (X 0TXT )
1 ΥT
01
12
=T
32δ δ
s2T0 1
Υ1T X
0TXTΥ1T
1 01
12
=T
32δ δ
24s2T 0 1 " T
T 3T (T+1)2T 3
T (T+1)2T 3
T (T+1)(2T+1)6T 3
#1 01
35 12
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model V
=) 1sσ20 1
1 12
12
13
1 01
N 0, 12σ2
=1s
σ20 1
4 66 12
01
N 0, 12σ2
=1p12σ2
N0, 12σ2
= N (0, 1)
The result is that even though the parameter converges faster thatususal, we can still use our standard t test.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model VI
DenitionWhen a parameter estimate converges to its true value at a rate fasterthan
pT , we say that the estimate is super-consistent.
Now we will consider a single hypothesis involving both parameters.We wish to test:
H0 : r1α+ r2δ = r
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model VIIGenerally we would test hypotheses like this one using the square rootof the F test. This implies that
tT =
r1 α+ r2 δ r
s2Tr1 r2
(X 0TXT )
1r1r2
12
=
r1 α+ r2 δ r
s2Tr1 r2
Υ1T ΥT (X 0TXT )
1 ΥTΥ1T
r1r2
12
=
r1 α+ r2 δ r
"s2ThT
12 r1 T
32 r2
i Υ1T X
0TXTΥ1T
1 " T 12 r1
T32 r2
## 12
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model VIII
tT =
r1 α+ r2 δ r
"s2ThT
12 r1 T
32 r2
i Υ1T X
0TXTΥ1T
1 " T 12 r1
T32 r2
## 12
=
pTr1 α+ r2 δ r
s2Tr1 T1r2
Υ1T X
0TXTΥ1T
1 r1T1r2
12
Now, under the null hypothesis,pTr1 α+ r2 δ r
=
pTr1 α+ r2 δ (r1α+ r2δ)
= r1
pT (α α) + r2
pTδ δ
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model IX
Now,
pTδ δ
P! 0
andpT (α α)) N
0, 4σ2
This implies that
pTr1 α+ r2 δ r
) r1N
0, 4σ2
= N
0, 4r21 σ2
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model XTo nd the limit of the demonator, note that
s2TP! σ2r1 T1r2
!r1 0
and
Υ1T X0TXTΥ1T
1 ! 4 66 12
implying that
s2Tr1 T1r2
Υ1T X
0TXTΥ1T
1 r1T1r2
! σ2
r1 0
4 66 12
r10
= σ24r21
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model XI
such that
tT =
pTr1 α+ r2 δ r
s2Tr1 T1r2
Υ1T X
0TXTΥ1T
1 r1T1r2
12
=) 1σ24r21
N0, 4r21 σ2
= N (0, 1)
As a result we can, again, use the standard t test for this type ofhypotheses.
Note that this test has the same distribution under the null as itwould have had if we used the true value of δ instead of an estimate.This is due to the super-consistency. The slowest convergingparameter estimate dominates the asymptotic distribution.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model XIINow consider a joint test of separate hypotheses about α and δ :
H0 :
αδ
=
α0δ0
or
H0 : β = β0
For this we use a Wald test. The standard test statistic for thishypothesis is:
W =
β β00 hs2TX 0TXT
1i1 β β0
=
β β0
0ΥThs2TΥT
X 0TXT
1 ΥTi1
ΥT
β β0
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model XIIIBecause
s2TΥTX 0TXT
1 ΥT ! σ24 66 12
and
ΥT
β β0) N
0,4 66 12
σ2
we have
W =
β β00
ΥThs2TΥT
X 0TXT
1 ΥTi1
ΥT
β β0) χ2 (2)
We have thus established that for the simple regression model with adeterministic trend, all the usual test statistics are valid.
Helle Bunzel ISU
Non-stationary Time Series
Introduction Comparison of Unit Root and Trend-Stationary Series Asymptotic Theory Trend-Stationary Series
Testing in the Simple Trend Model XIV
The rest of Hamilton Chapter 16 is spent on more complicated trendmodels with ARMA errors. The proofs are more complicated, but theend results are the same: We can apply the usual asymptotic testsand the have the usual asymptotic distributions.
Helle Bunzel ISU
Non-stationary Time Series