Time Series Econometrics 2011

Helmut Ltkepohl

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

Part I Univariate Time Series

1. 2. 3. 4. 5. 6. Stationary and integrated stochastic processes ARIMA processes Estimation and specification of stationary ARMA processes Forecasting Estimation of I(1) processes and unit root tests Spectral Analysis / Frequency Domain Analysis

Part II
7. Dynamic regression models: setup and estimation

Part III Multiple Time Series

8. 9. 10. 11. Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly changes in U.S. fixed investment

quarterly German long-term interest rate

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly German nominal GNP

daily log DAFOX

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Stochastic processes
, y1 , y2 , y3 , y1 , y 2 , ,yT
sequence of random variables (stochastic process) observations (time series)

Notation: or

yt
yt

for stochastic process for time series

sometimes

y1 , ..., y T
Time Series Econometrics 2011

Helmut Ltkepohl

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Stationary stochastic processes

yt is stationary if

(1) E ( yt ) = y for all t (2) E ( yt y )( yt h y ) = h for all t and h

(the first and second moments of

yt are time invariant)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly German nominal GNP

Is this series stationary (generated by a stationary process) ?

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly changes in U.S. fixed investment

Is this series stationary ?

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Estimation of moments
yt
stationary stochastic process,

y 1 , .. . , y T

time series (sample)

1 T y = y = yt T t =1

(sample mean) estimator for

1 T h = ( yt y )( yt h y ) T t = h +1 1 T h = 1 ( yt y )( yt h y ) T h t =h+

(sample autocovariances) estimators of h

h = h / 0
h = h / 0

(sample autocorrelations) estimators of autocorrelations

h = h / 0

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

White noise (WN)

yt
is a stationary stochastic process with

E ( yt ) = 0

h = 0 for h = 1, 2,

yt

is called white noise

d T h N ( 0,1)

or or

1 h N ( 0, T )

T ,2

is approximate 95% confidence interval around zero

Time Series Econometrics 2011

Helmut Ltkepohl

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly changes in U.S. fixed investment

autocorrelations of investment

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly German long-term interest rate

autocorrelations of long-term interest rate

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Partial autocorrelations
yt stationary stochastic process
ah = Corr ( yt , yt h yt 1 , , yt h +1 )
h-th partial autocorrelation coefficient estimator of ah :

ah OLS estimator of h in

yt = + 1 yt 1 + + h yt h + ut

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly changes in U.S. fixed investment

autocorrelations of investment

partial autocorrelations of investment

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly German long-term interest rate

autocorrelations of long-term interest rate

partial autocorrelations of long-term interest rate

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Transformations
income log income

yt

log income = yt yt 1

4 log income = yt yt 4

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Integrated processes
yt nonstationary stochastic process
yt := yt yt 1 stationary yt ~ I (1) (integrated of order 1) yt nonstationary but

2 yt = ( yt ) = yt 2 yt 1 + yt 2 stationary yt ~ I (2) etc. yt stationary yt ~ I (0)

yt quarterly nonstationary process 4 yt = yt yt 4 stationary yt seasonally integrated

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Autoregressive processes
AR(1)
yt = yt 1 + ut , = i ut i
i =0

ut ~ WN 0, u2 if < 1

moving average (MA) representation more formally

L lag operator, that is, Lyt = yt 1

(1 L ) yt = ut

yt = (1 L )

i i ut = L ut i =0

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

y = E ( yt ) = i E (ut i ) = 0
i =0 i h = E ( yt yt h ) = E ut i j ut h j = u2 j + h j j =0 j =0 i =0

u2 h = 1 2

h = 0, 1, 2,

yt is stationary if < 1
or, equivalently, if

1 z = 0 holds for z > 1

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

AR(1)-Process,

=0.5

AR(1)-Process,

=-0.5

Autocorrelation Function (ACF)

Autocorrelation Function (ACF)

Partial Autocorrelation Function (PACF)

Partial Autocorrelation Function (PACF)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

AR(1) with nonzero mean

yt = + yt 1 + ut
or

(1 L ) yt = + ut
1 1 yt = + (1 L ) ut = + (1 L ) ut 1 L 1
E ( yt ) =

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Random Walk AR(1) with = 1

yt = yt 1 + ut , = y0 + ui
i =1 t 2 ut ~ WN 0, u

E ( yt ) = y0

for fixed y0 nonstationary

E ( yt y0 ) 2 = t u2 yt
However, yt = ut yt ~ I (1)

is stationary

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Random Walk with drift

yt = + yt 1 + ut = y0 + t + ui
i =1 t

E ( yt ) = y0 + t

yt has linear trend in mean yt = yt yt 1 is stationary

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

AR(p) process

yt = + 1 yt 1 + + p yt p + ut =

1 1 p

+ i ut i
i =0

if 1 1 z p z p 0 for z 1
alternatively

1 1 L p Lp yt = ( L) yt = + ut

yt = hence,

1 1 p
i i

+ 1 1 L p L
p

ut

L = (1 L
i =0 1
Helmut Ltkepohl

pL

= ( L )1

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

AR(p) process

(contd)

y = E ( yt ) =

1 1 p

h = E ( yt y )( yt h y ) = E i ut i j ut h j j =0 i =0 2 = h j + h j , j =0

h = 0, 1, 2,

yt is stationary if 1 1 z p z p 0 for z 1

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

AR(p) process

(contd)
1 z p

1 1 z p z p = 0 for z =1 , , p

1 1 z p z p = 1 z 1

1 1 z p z p = 1 1* z * 1 z p 1 (1 z ) if one i = 1 p
1 1* L * 1 Lp 1 (1 L) yt p = 1 1* L * 1 Lp 1 yt = + ut p

is stationary if 1 1* z * 1 z p 1 0 for z 1 p
yt ~ I (1)
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

MA(q) process

yt = + ut + m1ut 1 + + mq ut q = + 1 + m1 L + + mq Lq ut q = + m( L)ut is stationary yt is called invertible if m( z ) 0 for z 1 m( L) 1 yt = m(1) 1 + ut

or

yt = + i yt i + ut
i =1

(AR representation)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

MA(1)-Process,

m=0.8

MA(1)-Process, m=-0.8

Autocorrelation Function (ACF)

Autocorrelation Function (ACF)

Partial Autocorrelation Function (PACF)

Partial Autocorrelation Function (PACF)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

ARMA(p,q) process
yt = + 1 yt 1 +
or with

+ p yt p + ut + m1ut -1 +

+ mq ut q

( L) yt = + m( L)ut ( z ) 0 for z 1
m( z ) 0 for z 1

is stationary and invertible

+ ( L) 1 m( L)ut (1) has AR representation m( L) 1 ( L) yt = + ut

has MA representation yt =

m(1)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

ARIMA(p,d,q) process
( L) d yt = + m( L)ut
is I (d )

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Estimation of AR models
AR(p)
yt = + 1 yt 1 + ... + p yt p + ut , ut ~ WN 0, u2

Estimate , 1 ,..., p by OLS.

If yt is stationary, the estimators have the usual asymptotic properties. E.g., AR(4) investment series

yt = 0.82 + 0.51 yt 1 0.10 yt 2 + 0.06 yt 3 0.22 yt 4 + ut

(2.76) (4.86) (-0.83) (0.54) (-2.02)
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Estimation of ARMA models

ARMA(p,q)

( L) yt = m( L)ut ,
ML estimation
log-likelihood

ut ~ WN 0, u2

normally distributed

l (1 ,..., p , m1 ,..., mq ) = lt ()
t =1

with

1 1 lt () = log 2 log u2 m( L) 1 ( L) yt 2 2

2 u2

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Model specification
Specification of AR order Estimate AR(n) model for Choose

n = 0,..., pmax

p such that it minimizes a criterion such as

2 u

2 AIC (n) = log (n) + n T HQ(n) = log u2 (n) + SC (n) = log u2 (n) + 2 log log T n T log T n T

(Akaike) (Hannan-Quinn) (Schwarz, Rissanen)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Order selection criteria for U.S. investment series

n 0 AIC(n) 2.170 HQ(n) SC(n) 2.180 2.195 1 1.935 1.956 1.987 2 1.942 1.974 2.020 3 1.950 1.997 2.059 4 1.942 1.995 2.073 5 1.963 2.027 2.122 6 1.990 2.065 2.176 7 2.018 2.104 2.231 8 1.999 2.097 2.241 9 1.997 2.107 2.268 10 2.032 2.153 2.331

p ( SC ) p( HQ) p( AIC )

(if T 16)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Specification of ARMA models

Box-Jenkins approach based on sample autocorrelations and partial autocorrelations

Helmut Ltkepohl

Time Series Econometrics 2011

Autocorrelation Function (ACF)

Partial Autocorrelation Function (PACF)

Autocorrelations of AR(1)

Autocorrelations of AR(1)

Partial Autocorrelations of AR(1)

Partial Autocorrelations of AR(1)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Autocorrelations of AR(2)

Autocorrelations of AR(2)

Partial Autocorrelations of AR(2)

Partial Autocorrelations of AR(2)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Autocorrelations of ARMA(1,1)

Autocorrelations of ARMA(1,1)

Partial Autocorrelations of ARMA(1,1)

Partial Autocorrelations of ARMA(1,1)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Hannan-Rissanen approach

(i ) Fit long AR( h) with large h by OLS yt = 1 yt 1 + + h yt h + ut

to get residuals ut ( h)
(ii ) Choose ARMA orders by estimating yt = 1 yt 1 + + n yt n + ut + m1ut 1 (h) + + ml ut l (h)

and comparing models with different n, l.

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Model checking
AR(4) for investment series

Plot residuals and residual autocorrelations

standardized residuals 1948.21972.4, T=99

residual autocorrelations

residual partial autocorrelations

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Portmanteau test for residual autocorrelation

H 0 : u ,1 = u ,2 = ... = 0
vs.

where u ,i = Corr (ut , ut i )

H 1 : u ,i 0

for at least one i {1, 2,...}

test statistic

Qh = T u2, j
j =1

u , j

1 T s s = ut ut j T t = j +1

or
* h 2 h

uts standardized residual

1 Q =T u2, j 2 (h p q) j =1 T j
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

LM test for residual autocorrelation in AR models

Breusch-Godfrey test

ut = 1ut 1 +
H 0 : 1 =

+ h ut h + errort
vs.

= h = 0

H1 : 1 0 or ...or h 0

estimate auxiliary model

ut = + 1 yt 1 +
LM h = TR 2 2 (h)
or

+ p yt p + 1ut 1 +

+ hut h + et

R2 T p h 1 FLM h = F (h, T p h 1) 2 1 R h
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Other model checks

test for nonnormality ARCH test nonlinearity test (RESET) stability analysis: Chow tests, CUSUM tests, recursive analysis etc.

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Forecasting
DGP yt = 1 yt 1 +
+ p yt p + ut

1-step ahead forecast at forecast origin T:

yT +1 T = 1 yT +

+ p yT +1 p

h-step ahead forecast at forecast origin T:

yT + h T = 1 yT + h 1 T + + p yT + h p T

(compute forecasts recursively for h=1,2,)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

forecast error
yT + h yT + h T = uT + h + 1uT + h 1 +
2 + h 1uT +1 ~ 0, y (h)

coefficients of MA representation
2 y 2 2 u

(h) = E ( yT + h yT + h T ) =
ut ~ N (0, u2 )

j2
j =0

h 1

95% forecast interval yT + h T 1.96 y (h), yT + h T + 1.96 y (h)

in practice: replace unknown parameters by estimates all results hold asymptotically or approximately under suitable assumptions
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Unit root tests

Augmented Dickey-Fuller (ADF) test DGP: yt = 1 yt 1 +

+ p yt p + ut p = 0

has unit root if 1 1

Reparameterize process:

yt = yt 1 + 1*yt 1 + where = (1 1
test

+ * 1yt p +1 + ut p p ),

* = ( j +1 + j

+p)

H 0 : = 0 vs. H1 : < 0
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

estimate model by OLS test statistic:

t = t -statistic for

reject H 0 if t is small critical values depend on deterministic terms in the model

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

KPSS test
H 0 : yt ~ I (0)

(Kwiatkowski, Phillips, Schmidt, Shin)

vs. H1 : yt ~ I (1)

DGP: yt = xt + zt

xt = xt 1 + vt
2 test H 0 : v =0 vs.

RW

H1 : v2 > 0

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

1 KPSS = 2 T
t j =1

2 St2
t =1

St = ( y j y )

2

T is estimator of = lim T Var zt T t =1

2 1

e.g.,

lq 1 T 1 T 2 = ( yt y ) + 2 j ( yt y )( yt j y ) T t =1 j =1 T t = j +1

where j = 1
e.g.,

j and lq lag truncation parameter lq + 1

1 4

lq = q (T /100 )

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Other unit root tests

higher order integration (test also yt or 2yt) structural breaks seasonal unit roots

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Spectral Analysis / Frequency Domain Analysis

Let a and h be independent random variables such that
2 a ~ (0, a ) and h ~ U (0, )

Define

yt = a cos( t + h) E ( yt ) = E (a ) E [ cos( t + h) ] = 0

Cov( yt , yt + j ) = E ( yt yt + j ) =

2 a

cos( j )

yt is stationary
Note: yt is deterministic
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Terminology:

yt = a cos( t + h)

amplitude frequency (number of cycles per unit measured in radians) phase wavelength, period

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

More generally:
a1 ,..., aM , a j ~ (0, 2 ), j h1 ,..., hM ~ U (0, )

mutually independent random variables

yt = ak cos(k t + hk )
k =1

E ( yt ) = 0 Cov( yt , yt + j ) =
M
2 k

k =1

cos(k j )

yt is stationary
but deterministic

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Judge et al. (1985)

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Alternative representation of yt

cos( A + B ) = cos( A) cos( B ) sin( A) sin( B) thus, ak cos(k t + hk ) = ak cos(k t ) cos(hk ) ak sin(k t ) sin( hk )

Define k = ak cos( hk ), k = ak sin( hk )

yt = [ k cos(k t ) + k sin(k t ) ]
k =1

stationary, deterministic process

Note that

h ~ U (0, ) Cov [ cos(h),sin(h) ] = E [ cos(h) sin( h) ] = 0

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Spectral Representation Theorem

Let yt be a zero mean stationary stochastic process with autocovariances

j , j = 0, 1, 2,..., such that

j
j =0

<,

then there exist random variables ( ), ( ) with zero mean such that

yt =

0 ( ) cos( t ) + ( ) sin( t ) d

4
3

and for any 0 < 1 < 2 < 3 < 4 < ,

2
1

( ) d and ( ) d and ( ) d and

( ) d are uncorrelated ( ) d are uncorrelated ( ) d are uncorrelated

Time Series Econometrics 2011

2
1

4
3

and for any 0 < 1 < 2 < and 0 < 3 < 4 <

2
1

4
3

Helmut Ltkepohl

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Spectral Density / Power Spectrum

The function f y : [ ,

defined as

1 f y ( ) = 2
of yt . Note that

0 + 2 j cos( j ) j =1

is called spectral density function or power spectrum or spectrum

ei = cos + i sin 2 cos = ei + e i

1 f y ( ) = 2 1 = 2
i j i j + e ) 0 + j (e j =1

j =

j e i j
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Properties
(a) Symmetry around zero:

f y ( ) = f y ( )

(b)

ei j f y ( ) d = j

in particular
2 f y ( ) d = 0 = y

f y ( )

represents the contribution of cycles of frequency to the total variability of yt .

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

(c)

2 2 ut ~ iid(0, u ) fu ( ) = u 2

(d) yt =

j xt j = ( L) xt
j =0
2

f y ( ) = ( e i ) ( e i ) f x ( ) = ( e i )
in particular

f x ( )

yt ~ ARMA( p, q ),
2

( L) yt = m( L)ut

2 m ( ei ) u f y ( ) = (ei ) 2

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

2 1 u f y ( ) = 2 1 ei 1

2 1 u = 2 1 + 12 21 cos

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

2 f y ( ) = (2 ) 1 u 1 + m1ei

2 2 = (2 ) 1 u (1 + m1 + 2mi cos )

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Estimation of the Spectrum

1 f y ( ) = 2
0 + 2 j cos( j ) j =1

Periodogram

1 I y ( ) = 2

T 1 0 + 2 j cos( j ) j =1

inconsistent estimator of f y ( ) !

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Better

1 f y ( ) = 2

MT 0 0 + 2 j j cos( j ) j =1

where j ( j = 1,..., M T ) are the weights of the spectral window with window width MT.

Bartlett window: j j = 1 MT

ensures consistency of f y ( ) if M T and M T T 0 as T .

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly growth rates of German consumption

periodogram

MT=24

MT=12

MT=6

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly German inflation

periodogram

log periodogram

MT=20

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

quarterly German log income

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

1. Stationary and integrated stochastic processes 2. ARIMA processes 3. Estimation and specification of stationary ARMA processes 4. Forecasting 5. Estimation of I(1) processes and unit root tests 6. Spectral Analysis / Frequency Domain Analysis

Filtering: removal or extraction of components related to specific frequencies

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

Dynamic regression models

Example

yt = 1 yt 1 + + 0 xt + 1 xt 1 + ut
Estimation by OLS if

ut

is WN and uncorrelated

with xt and xt-1.

Standard asymptotic properties under usual assumptions! IV estimation if xt and ut are correlated or if ut is autocorrelated. Avoid autocorrelated ut by adding further lags!

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

Interpretation (1 1 L) yt = + ( 0 + 1 L ) xt + ut
yt =

+ 1 L 1 xt + ut + 0 1 1 1 1 L 1 1 L i i = + 1 L ( 0 + 1 L ) xt + wt 1 1 i =0 i i 1 = + 0 xt + (1 0 + 1 1 )xt i + wt = + i xt i + wt 1 1 i =0 i =1

distributed lag model with impact multiplier or short run multiplier 0 (= 0 ) cummulated effect

0 + 1 long-run multiplier or long-run effect i = 1 1 i =0 exists only if 1 1.

i= 0

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

More general model

yt = 1 yt 1 +
or or

+ p yt p + + 0 xt +

+ q xt q + ut

(1 1 L

p Lp ) yt = + ( 0 +

+ q Lq ) xt + ut

0 + + q Lq 1 yt = + xt + u p p t 1 1 p 1 1 L p L 1 1 L p L
if 1 1 z pz p 0 for z 1

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

Spurious regression problem

yt = + xt + ut yt and xt independent RWs (i.e., =0)
OLS regression
2 t for T

reject H 0 : = 0 with probability approaching 1 as T

possible solution:

use yt = + xt + vt use lags of yt !

Time Series Econometrics 2011

Problem: cointegration! alternative solution:

Helmut Ltkepohl

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Vector autoregressive models

0.1

0.2

investment investment

22

42

62

82

3000

0.15

investment investment

2000

0.03

consumption consumption

0.05

income income

income income

22

42

62

82

1000

0.03

1960.1

1965.1

1970.1

0.02

1975.1

1980.1

0.05

consumption consumption

22

42

62

82

Helmut Ltkepohl

Time Series Econometrics 2011

0.02

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

VAR(p)
yt = ( y1t ,..., yKt ) ' K -dimensional

yt = A1 yt 1 +

+ Ap yt p + ut

Ai (KxK) parameter matrix ut ~ WN ( 0, u ) , u positive definite

stationarity (stability) condition

det( I K A1 z

Ap z p ) 0 for z 1
Time Series Econometrics 2011

Helmut Ltkepohl

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

MA representation of stable VAR(p)

yt = I K A1 L

Ap L

ut = j ut j
j =0

E ( yt ) = 0 yt = + A1 yt 1 + = ( I K A1 + Ap yt p + ut Ap ) + j ut j
1
j =0

E ( yt ) = ( I K A1
Helmut Ltkepohl

Ap )
1
Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Forecasting ( yt = A1 yt 1 +
1-step ahead forecast at origin

+ Ap yt p + ut )

yT +1 T = A1 yT +

+ Ap yT p +1

h-step ahead forecast at origin T yT + h T = A1 yT + h 1 T + + Ap yT + h p T

forecast error

yT + h yT + h T = j uT + h j ~ ( 0, y (h) )
j =0 h 1

h 1

with forecast error covariance or MSE matrix

y (h) = j u j
j =0
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Granger-causality
zt yt = xt t all relevant information in the universe at time t xt is Granger-causal for zt iff z ( h t ) z h t \ { xs s t }

for at least one h = 1, 2,...

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Granger-causality in VAR
zt p A11,i yt = = xt i =1 A21,i A12,i zt i u zt + A22,i xt i u xt

xt is not Granger-causal for zt iff

A12,i = 0 for i = 1,..., p

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Instantaneous causality
xt is instantaneously causal for zt iff z 1 t { xt +1} z (1 t ) zt x VAR(p) t xt is not instantaneously causal for zt iff E ( u zt u ) = 0 xt zt is not instantaneously causal for xt
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Impulse response analysis

yt = ut + 1 ut 1 + 2 ut 2 +
elements of

is: forecast error impulse responses

Example

y1,t .5 0 0 y1,t 1 u1,t y2,t = .1 .1 .3 y2,t 1 + u2,t y3,t 0 .2 .3 y3,t 1 u3,t

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Structural impulse responses

t = B 1ut ~ ( 0, I K )
yt = ut + 1 ut 1 + 2 ut 2 + = B t + 1 t 1 + 2 t 2 +

with i = i B

orthogonalized or structural impulses

often:

0 b11 b 21 b22 B= bK 1 bK 2
Helmut Ltkepohl

0 0 bKK

triangular

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Forecast error variance decomposition

yT + h yT + h T = j uT + h j = j T + h j
j =0 j =0 h 1 h 1

h 1 ~ 0, j j j =0

yk ,T + h yk ,T + h T = ( k1, j 1,T + h j +
j =0 K

h 1

+ kK , j K ,T + h j ) + ki ,h 1 i ,T +1 )
2 + ki ,h 1 ) 2

= ( ki ,0 i ,T + h +
i =1

E yk ,T + h yk ,T + h T

2 = ( ki ,0 + K i =1

2 ki ,0

2 ki , h 1

E yk ,T + h yk ,T + h T

is the contribution of innovation error variance of variable k

Helmut Ltkepohl

i to h-step forecast

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Forecast error variance decomposition of the investment/income/consumption system

forecast error in investment forecast horizon h 1 2 3 4 5 10 1 2 3 4 5 10 1 2 3 4 5 10 proportions of forecast error variance h periods ahead accounted for by innovations in investment income consumption 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 1 0 0 0 1 0 .020 .941 .039 .026 .930 .044 .029 .926 .045 .030 .925 .045 .030 .925 .045 .030 .925 .045 0 .338 .662 0 .411 .589 .001 .421 .578 .002 .423 .576 .002 .423 .575 .002 .423 .575 .002 .423 .575
Time Series Econometrics 2011

income

consumption

Helmut Ltkepohl

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Estimation of VAR(p)
yt = + A1 yt 1 + + Ap yt p + ut

estimate parameters by equationwise OLS

ut ~ N ( 0, u ) OLS = ML

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

OLS estimation of VAR(p)

1 yt 1 + ut yt = , A1,..., Ap yt p A
Zt 1

[ y1 ,..., yT ] = A [ Z 0 ,..., ZT 1 ] + [u1 ,..., uT ]

OLS estimator of

Y = A Z +U

1 A = Y Z ( Z Z )

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Asymptotic properties of OLS estimator

d T vec A A N ( 0, A )

A nonsingular if yt is stationary
standard inference possible (e.g. for causality analysis)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Bayesian Estimation
Parameter vector of interest Prior probability density g ( ) Sample probability density f ( y1 ,..., yT | ) Posterior probability density

g ( | y1 ,..., yT ) = f ( y1 ,..., yT | ) g ( ) / f ( y1 ,..., yT )

(Bayes Theorem)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

y1t = 11,1 y1,t 1 + 12,1 y2,t 1 + 11,2 y1,t 2 + 12,2 y2,t 2 + u1t ,
( ) ( 1 / 2 ) ( / 2) ( 1 / 2 2 )

Litterman or Minnesota Prior

y2t = 21,1 y1,t 1 + 22,1 y2,t 1 + 21,2 y1,t 2 + 22,2 y2,t 2 + u2t ,
( 2 / 1 )
=
2

( )

( 2 / 2 1 )

( / 2)
2 2

2 2 1

1 2 2

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Estimation of impulse responses

= ( A1 ,..., Ap , u ) vector of impulse responses
= ( A1 ,..., Ap , u ) estimator

d T ( ) N (0, )

(by delta method)

where = with

= vec( A1 ,..., Ap , u )

can be used to construct confidence intervals for impulse responses

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Residual based bootstrap

(1) Estimate

yt = + A1 yt 1 + + Ap yt p + ut to get residuals u1 ,..., u T .

(2) Obtain bootstrap residuals

* * u1 ,..., uT by randomly drawing residual vectors from u1 ,..., u T with replacement.

(3) Compute bootstrap time series

yt* = + A1 yt*1 +

+ Ap yt* p + ut* , t = 1,..., T

* A1* ,..., Ap , * u

* * with ( y p +1 ,..., y0 ) = ( y p +1 ,..., y0 ) .

(4) Obtain estimates (5) Obtain

from bootstrap time series.

* * = ( A1* ,..., Ap , * ) . u

(6) Repeat steps (2) (5)

times with

a large number.

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Bootstrap confidence intervals

Suppose is a single impulse response coefficient.

Standard percentile interval
* Let s* 2 and s1 2

be the ( 2 ) - and (1 2 ) -quantiles

* of the distribution 1* ,..., N and define

* CI S = s* 2 , s1 2 .

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Choose VAR order by criteria such as

AIC (n) = log u ( n ) +

2 (number of freely estimated parameters) T 2 = log u ( n ) + nK 2 T

2 log log T HQ(n) = log u ( n ) + nK 2 T

log T SC (n) = log u ( n ) + nK 2 T
fit subset models
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. 9. 10. 11.

Vector autoregressive models Estimation and specification of VAR models Cointegration and vector error correction models Estimation and specification of VECMs

Model checking
visual inspection of residuals tests for residual autocorrelation (Portmanteau and LM) tests for nonnormality ARCH tests stability analysis (CUSUM, Chow tests etc.) Forecasting and impulse response analysis with estimated processes: replace unknown quantities by estimates
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Cointegration

U Rt S :

U rt S : - - -

EU Rt :

EU rt : - - -

U U Rt S r t S

EU EU Rt r t

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

RUS, rUS, REU, rEU ~I(1) RUS rUS, REU rEU ~I(0) RUS and rUS REU and rEU
are cointegrated are cointegrated

more generally:

yt ~ I(1), c a fixed K-dimensional vector c'yt ~ I(0) yt is cointegrated

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

VAR(p)
is integrated if

yt = A1 yt 1 + + Ap yt p + ut

det( I K A1 z

Ap z p ) = 0 for z = 1

reparameterization as in ADF test

yt = yt 1 + 1 yt 1 + = ( I K A1 i = ( Ai +1 + Ap )

+ p 1 yt p +1 + ut

+ Ap ) , i = 1,..., p 1

yt ~ I (1) if rk ( ) = r < K
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

( K K ), rk( ) = r there exist , ( K r ), such that = yt = yt ~ I (0)

1

rk( ) = rk( ) = r ,

( ) yt = yt ~ I (0)

yt

are the cointegration relations are not unique; take any nonsingular (rr) matrix

and

and define

* = Q, * = Q 1 * * = =

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Vector error correction model (VECM)

yt = yt 1 + 1 yt 1 + ... + p 1 yt p +1 + ut
short-term dynamics

, ( K r ), r =

cointegration rank

yt 1 yt

error correction term cointegration or long-run relations

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Example 1

RtUS 1 RtUS RtUS = [1, 1] US1 + 1 US1 + US rt 2 rt 1 rt 1

Example 2

+ ut

y1t ~ I (1) y2t ~ I (0)

y1t yt = ~ I (1) y2 t + ut

y1,t 1 1 yt = [ 0, 1] + 1 yt 1 + 2 y2,t 1
Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

cointegration rank r can be chosen such that

Ir = ( K r )
Example 3

r =1 yt = (1, 2 ,..., K ) yt 1 + 1 yt 1 + + ut

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Example 4
11 y t = 21 31

12 1 0 22 0 1 32

1 y t 1 + 1 y t 1 + 2

+ ut

every pair of components of yt is cointegrated cointegration can be explored in bivariate subsystems of yt Attention:

1 and 2 may be zero !

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Deterministic terms
yt = t + xt , E ( xt ) = 0, E ( yt ) = t
+ p 1 xt p +1 + ut xt = xt 1 + 1 xt 1 +

t =
xt = yt , xt = yt yt = ( yt 1 ) + 1 yt 1 + y t 1 + 1 yt 1 + = 1
0
Helmut Ltkepohl

+ p 1 yt p +1 + ut + ut

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Linear trend
t = 0 + 1t
xt = yt 0 1t , xt = yt 1
yt 1 = ( yt 1 0 1 ( t 1) ) + 1 ( yt 1 1 ) + + ut

y t 1 + 1 yt 1 + yt = + t 1
+

+ ut if 1 = 0

or yt = + yt 1 + 1 yt 1 +

+ ut

(i.e., the trend is orthogonal to the cointegration relations)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Forecasting
convert VECM to VAR(p) representation

yt = A1 yt 1 +

+ Ap yt p + ut

h-step ahead forecast at origin T yT + h T = A1 yT + h 1 T + + Ap yT + h p T

forecast error:

yT + h yT + h T = j uT + h j ~ ( 0, y (h) )
j =0

h 1

y (h) = j u j
j =0

h 1

where the

i are computed from the Ai

stationary case: I(1) case:

y (h) y for h y (h) unbounded for h

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Granger-causality in cointegrated VAR

zt p A11,i yt = = xt i =1 A21,i A12,i zt i u zt + A22,i xt i u xt

xt is not Granger-causal for zt iff

A12,i = 0 for i = 1,..., p
as in stationary VAR

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Impulse response analysis

MA representation

yt = ut + 1 ut 1 + 2 ut 2 + = B t + 1 t 1 + 2 t 2 +
does not exist in I(1) case. Still, impulse responses can be computed from same way as in I(0) case.

with i = i B Ais in the

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Estimation of cointegrated VAR

yt = A1 yt 1 + + Ap yt p + ut
equationwise OLS estimator:
d T vec A A N ( 0, A )

A = A1,..., Ap

Attention:

A is singular in general
F tests may not be valid tests for Granger-causality may be invalid

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Test for Granger-causality

zt p A11,i yt = = xt i =1 A21,i A12,i zt i u zt + A22,i xt i u xt

H 0 : A12,i = 0 for i = 1,..., p

fit

yt = A1 yt 1 +

+ Ap yt p + Ap +1 yt p 1 + ut

and test

H0 with usual 2 (Wald) test!

(lag augmentation test)

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Estimation of VECM
yt 1 +u yt = yt 1 + 1 ,..., p 1 t yt p +1
X t 1

Y = [ y1 ,..., yT ] X = [ X 0 ,..., X T 1 ]

Y1 = [ y0 ,..., yT 1 ] U = [u1 ,..., uT ]

Y = Y1 + X + U
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

, known

= ( Y Y ) X ( X X )1 1
1

Y = Y1 + ( Y Y1 ) X ( X X ) X + U *
or with

R0 = R1 + U * R0 = Y IT X ( X X ) X
1

R1 = Y1
estimate

( (I

X ( X X )

) X )

from this model

e.g., by reduced rank regression or Gaussian ML (Johansen procedure)

Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

If

is normalized such that

Ir = ( K r )

then T ( K r ) ( K r ) converges in distribution

Attention:

can only be interpreted properly if identified (made unique)

standard inference is possible for and for

(if identified)

and

= Y Y X ( X X )1 1
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Specification of cointegrating rank

test sequentially

H 0 : rk ( ) = 0 H 0 : rk ( ) = 1 H 0 : rk ( ) = K 1

vs. vs. vs.

H1 : rk ( ) > 0 H1 : rk ( ) > 1 H1 : rk ( ) = K

use Johansens LR tests or other systems cointegration tests choose

r such that H 0 : rk ( ) = r is the first null hypothesis

which cannot be rejected

Helmut Ltkepohl

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

The critical values of the tests depend on the deterministic terms in the VECM Main cases:

(1)

yt = + xt
0

(constant)

yt 1 yt = + 1 yt 1 + + ut 1 (2) yt = 0 + 1t + xt (unrestricted linear trend) yt 1 yt = + + 1 yt 1 + + ut t 1

+

(3)

yt = 0 + 1t + xt , 1 = 0

yt = + yt 1 + 1 yt 1 + + ut
Helmut Ltkepohl

(linear trend in variables but not in cointegration relations)

Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Specification of VECM
y1t,,yKt time series of interest
Step 1: Determine orders of integration of ykts Step 2: Determine cointegration relations between all integrated variables, also considering subsystems of variables Step 3: Set up, estimate and simplify an overall VECM for yt=( y1t,,yKt ) Step 4: Model checking plot residuals test for residual autocorrelation test for nonnormality test for ARCH stability analysis
Helmut Ltkepohl Time Series Econometrics 2011

Part I Univariate Time Series Part II Dynamic regression models Part III Multiple Time Series

8. Vector autoregressive models 9. Estimation and specification of VAR models 10. Cointegration and vector error correction models 11. Estimation and specification of VECMs

Extensions
structural breaks in deterministic term I(2) variables seasonal cointegration

Helmut Ltkepohl

Time Series Econometrics 2011