Econometrics
J. Ortega Garcia
Deterministic and stochastic trend: an introduction1
1
This presentation is based on Hamilton(1994), Time series analysis
J. Ortega Garcia Econometrics 2016 1 / 15
1 ARMA model
2 Modeling nonstationary time series
General Form
Deterministic trend
Stochastic trend
3 Comparisons of trend-stationary and unit root processes
Forecast
Dynamic multipliers
Transformations to achieve stationarity
J. Ortega Garcia Econometrics 2016 2 / 15
ARMA model
1 ARMA model
2 Modeling nonstationary time series
General Form
Deterministic trend
Stochastic trend
3 Comparisons of trend-stationary and unit root processes
Forecast
Dynamic multipliers
Transformations to achieve stationarity
J. Ortega Garcia Econometrics 2016 3 / 15
ARMA model
ARMA model
A univariate time series model can be written in the form
yt = + t + 1 t1 + 2 t2 +
= + (L)t
2
where: t i.i.d N 0,
P
j=0 |j | <
roots of (L) = 0 are outside the unit circle
Two important characteristics:
E (yt ) = 0
lm yt+s|t =
s
Many of the economic and financial time series encountered in practice are
nonstationary.
J. Ortega Garcia Econometrics 2016 4 / 15
Modeling
1 ARMA model
2 Modeling nonstationary time series
General Form
Deterministic trend
Stochastic trend
3 Comparisons of trend-stationary and unit root processes
Forecast
Dynamic multipliers
Transformations to achieve stationarity
J. Ortega Garcia Econometrics 2016 5 / 15
Modeling General Form
Modeling nonstationary time series
General form
yt = + t + t
1 1 L 2 L p Lp t = 1 + 1 L + 2 L2 + + q Lq t
2
Suppose that autoregressive operator is factored:
1 1 L 2 L2 p Lp = (1 1 ) (1 2 ) 1 p
If the eigenvalues are inside the unit circle
yt = + t + (L)t (1)
1 + 1 L + 2 L2 + + q Lq t
t = = (L)t
(1 1 L) (1 2 L) 1 p L
P
with j=0 |j | < 0 and roots of (L) outside the unit circle.
J. Ortega Garcia Econometrics 2016 6 / 15
Modeling Deterministic trend
Modeling nonstationary time series
Deterministic trend
In (1), when |i | < 1 for i = 1, 2, , p the process is known as deterministic trend or
stationary in trend.
J. Ortega Garcia Econometrics 2016 7 / 15
Modeling Stochastic trend
Modeling nonstationary time series
Stochastic trend
Suppose instead that 1 = 1 and |i | < 1 for i = 2, 3, , p, then (1) would state that
1 + 1 L + 2 L2 + + q Lq
t = = (1 L)1 (L)t
(1 L) (1 2 L) 1 p L
yt = + t + (1 L)1 (L)t (2)
P
where j=0 |j | <
If (2) is first-differenced, the result is:
(1 L)yt = + (L)t
wich is the form of the unit root process.
J. Ortega Garcia Econometrics 2016 8 / 15
Comparisons of trend-stationary and unit root processes
1 ARMA model
2 Modeling nonstationary time series
General Form
Deterministic trend
Stochastic trend
3 Comparisons of trend-stationary and unit root processes
Forecast
Dynamic multipliers
Transformations to achieve stationarity
J. Ortega Garcia Econometrics 2016 9 / 15
Comparisons of trend-stationary and unit root processes Forecast
Comparisons of trend-stationary and unit root processes
Forecast
Note that (1) implies
yt+s = + (t + s) + 0 t+s + 1 t+s1 + 2 t+s2 +
The optimal linear forescast, error associated with this forecast and mean squared error
(MSE) associated, respectively.
Et [yt+s ] = + (t + s) + s t + s+1 t1 + s+2 t2 +
yt+s Et [yt+s ] = 0 t+s + 1 t+s1 + 2 t+s2 + + s1 t+1
E (yt+s Et (yt+s ))2 = 02 + 12 + 22 + + s1
2
2
Where Et denotes the expectative operator conditioned on t , t1 , t1 ,
The MSE increases
P with the forecasting
horizon s though given the assumption of absolute
summability j=0 |j | < the added uncertainty, when s is large, becomes negligible
lm E (yt+s Et (yy+s ))2 = 02 + 12 + 22 + 2
s
J. Ortega Garcia Econometrics 2016 10 / 15
Comparisons of trend-stationary and unit root processes Forecast
Comparisons of trend-stationary and unit root processes
Forecast
Solving unit root
By contrast, for a unit root process (2) the s-period-ahead forecast, error associated with
this forecast and mean squared error (MSE) associated, respectively .
Et (yt+s ) = s + yt + (1 + 2 + + s ) t + 2 + 3 + + s+1 t1 +
yt+s Et yt+s = 0 t+s + (0 + 1 ) t+s1 + + 0 + 1 + + s1 t+1
2 h 2 2 i 2
E yt+s Et yt+s = 0 + (0 + 1 )2 + + 0 + 1 + + s1
The MSE of a unit root process does not converge to any fixed value when s goes to infinity.
J. Ortega Garcia Econometrics 2016 11 / 15
Comparisons of trend-stationary and unit root processes Dynamic multipliers
Comparisons of trend-stationary and unit root processes
Dynamic multipliers
Deterministic trend
yt = + t + 0 t + 1 t1 +
yt+s = + (t + s) + 0 t+s + 1 t+s1 +
yt+s
= s
t
yt+s X
lm = 0, given |j | <
s t
j=0
Stochastic trend
yt+s yt
= + 1 + 2 + + s
t t
= 1 + 1 + 2 + + s
= (1)
J. Ortega Garcia Econometrics 2016 12 / 15
Comparisons of trend-stationary and unit root processes Transformations to achieve stationarity
Comparisons of trend-stationary and unit root processes
Transformations to achieve stationarity
Deterministic trend
Substract + t from yt to produce a stationary representation.
Stochastic trend
Difference the series
J. Ortega Garcia Econometrics 2016 13 / 15
Comparisons of trend-stationary and unit root processes Transformations to achieve stationarity
Appendix
Why logs?
Deterministic trend
yt = e+t+t , t N 0, 2
log (yt ) = + t + t
Stochastic trend
yt
(1 L) log(yt ) = log
yt1
yt yt1
= log +1
yt1
yt yt1
= Taylors expansion of first order
yt1
The asumption is that the rate of growth of the series is a stationary stochastic process.
J. Ortega Garcia Econometrics 2016 14 / 15
Comparisons of trend-stationary and unit root processes Transformations to achieve stationarity
Appendix
Solving unit root
yt+1 = yt + + 0 t+1 + 1 t +
yt+2 = yt+1 + + 0 t+2 + 1 t+1 +
.. ..
. .
yt+s = yt+s1 + + 0 t+s + 1 t+s1 +
yt+s = yt + s + 0 t+s + (0 + 1 ) t+s1 + (0 + 1 + 2 ) t+s2 + +
0 + 1 + + s1 t+1 + (1 + 2 + + s ) t + 2 + 3 + + s+1 t1 +
Return
J. Ortega Garcia Econometrics 2016 15 / 15
