Forecasting

Past values of time series → Model → Forecast future values

MOOC Econometrics
Notation:
Lecture 6.3 on Time Series: yt : time series of interest (t = 1, . . . , n)
Specification and Estimation xt : time series possible explanatory factor (restrict to one)
Dick van Dijk, Philip Hans Franses, Christiaan Heij PYt−1 = {yt−1 , yt−2 , . . . , y1 }: past information on y at time t
PXt−1 = {xt−1 , xt−2 , . . . , x1 }
Univariate time series forecast model: ybt = F (PYt−1 )
Forecast model with explanatory factor: ybt = F (PYt−1 , PXt−1 )

Aim: Optimal use of past information to get best forecasts.

Wish: Forecast error εt = yt − ybt uncorrelated with past information.

erasmus.png
Lecture 6.3, Slide 2 of 13, Erasmus School of Economics

Univariate time series model Test question

Forecast: ybt = α + β1 yt−1 + . . . + βp yt−p .

Forecast: ybt = F (PYt−1 ) where PYt−1 = {yt−1 , yt−2 , . . . , y1 }.
Forecast error εt = yt − ybt uncorrelated with ys for all s < t.
Find forecast model F so that εt = yt − ybt uncorrelated with PYt−1 .
Test
Popular choice: F linear function of p past values: Show that εt is white noise, i.e., εt is uncorrelated with εs for all t 6= s.

ybt = α + β1 yt−1 + . . . + βp yt−p . Answer:

Without loss of generality, consider case s < t.

yt = ybt + εt = α + β1 yt−1 + . . . + βp yt−p + εt .
Pp
εs = ys − α − j=1 βj ys−j linear function of yr , r ≤ s < t.
AR(p) model, because εt is white noise.

εt is uncorrelated with yr for all r < t, so also uncorrelated with εs .

erasmus.png erasmus.png
Lecture 6.3, Slide 3 of 13, Erasmus School of Economics Lecture 6.3, Slide 4 of 13, Erasmus School of Economics
Estimation Time series model with explanatory factor

Forecast: ybt = F (PYt−1 , PXt−1 ).

Pp
Forecast error: εt = yt − α − j=1 βj yt−j . Find F such that εt = yt − ybt uncorrelated with PYt−1 and PXt−1 .

Pn 2
Minimize sum of squared forecast errors: t=p+1 εt . Popular choice: linear F :
ybt = α + β1 yt−1 + . . . + βp yt−p + γ1 xt−1 + . . . + γr xt−r .
OLS!
Pp Pr
yt = ybt + εt = α + j=1 βj yt−j + j=1 γj xt−j + εt .
Estimation of ARMA models: Maximum Likelihood.
Autoregressive Distributed Lag model: ADL(p, r ).

Pn 2
Estimation: minimize t=m+1 εt , where m = max(p, r ) → OLS!

erasmus.png erasmus.png
Lecture 6.3, Slide 5 of 13, Erasmus School of Economics Lecture 6.3, Slide 6 of 13, Erasmus School of Economics

Granger causality Consequences of non-stationarity

Two variables of interest: yt and xt . Regression assumption A2 not satisfied: regressors yt−j are random.

Make ADL model for each variable: Standard OLS t- and F -tests hold true in large enough samples
yt = α + pj=1 βj yt−j + rj=1 γj xt−j + εt .
P P provided all variables in equation are stationary.
P ∗ P∗
xt = α∗ + pj=1 βj∗ xt−j + rj=1 γj∗ yt−j + ε∗t . So: First test for non-stationarity before any estimation.

xt helps to predict yt if γj 6= 0 for some j AR(1): yt = α + βyt−1 + εt , test H0 : β = 1 against H1 : −1 < β < 1.
yt helps to predict xt if γj∗ 6= 0 for some j
Rewrite: ∆yt = yt − yt−1 = α + (β − 1)yt−1 + εt = α + ρyt−1 + εt
xt is Granger causal for yt if it helps to predict yt , where ρ = β − 1
whereas yt does not help to predict xt .
So: ∆yt = α + ρyt−1 + εt , test H0 : ρ = 0 against H1 : ρ < 0.
Test H0 : γj∗ = 0 for all j = 1, . . . , r ∗ by means of F -test.

erasmus.png
Reject H0 of non-stationarity if tρb < −2.9 (not conventional -1.65!).
erasmus.png
Note: Two ADL equations are estimated bySlide
Lecture 6.3, OLS7 of per equation.
13, Erasmus School of Economics Lecture 6.3, Slide 8 of 13, Erasmus School of Economics
Test question Augmented Dicky-Fuller test

Test Two types of test equations: with or without deterministic trend.

Rewrite the AR(2) model yt = α + β1 yt−1 + β2 yt−2 + εt as
∆yt = δ + ρyt−1 + γ∆yt−1 + εt , and express the parameters (δ, ρ, γ) in Test without deterministic trend if data no clear trend direction:
terms of (α, β1 , β2 ). ∆yt = α + ρyt−1 + γ1 ∆yt−1 + . . . + γL ∆yt−L + εt
Answer:
Reject H0 of non-stationarity if tρb < −2.9
∆yt = yt − yt−1
= α + (β1 − 1)yt−1 + β2 yt−2 + εt Test with deterministic trend if data clear trend direction:
= α + (β1 + β2 − 1)yt−1 − β2 yt−1 + β2 yt−2 + εt ∆yt = α + βt + ρyt−1 + γ1 ∆yt−1 + . . . + γL ∆yt−L + εt
= α + (β1 + β2 − 1)yt−1 − β2 (yt−1 − yt−2 ) + εt
Reject H0 of non-stationarity if tρb < −3.5
= α + (β1 + β2 − 1)yt−1 − β2 ∆yt−1 + εt

So: δ = α, ρ = β1 + β2 − 1, and γ = −β2 . Choice lag L: serial correlation check, or AIC/BIC (see Lecture 3).
erasmus.png erasmus.png
Lecture 6.3, Slide 9 of 13, Erasmus School of Economics Lecture 6.3, Slide 10 of 13, Erasmus School of Economics

Summary of Specification and Estimation Cointegration and error correction model

AR model for yt : xt and yt are cointegrated if both series are non-stationary,

but a linear combination (say yt − cxt ) is stationary.
Step 1: Perform ADF test on yt .
→ Non-stationarity rejected → model yt
→ Non-stationarity not rejected → take ∆yt yt = cxt : long-run equilibrium.
and perform ADF test on ∆yt
Step 2: Estimate AR model for stationary series by OLS. Engle-Granger test for cointegration:
→ Step 1: OLS in yt = α + βxt + εt → b and residuals et
ADL model for yt with explanatory factor xt : → Step 2: Cointegrated if ADF test on et rejects non-stationarity
∆et = α + ρet−1 + γ1 ∆et−1 + . . . + γL ∆et−L + ωt
Step 1: Perform ADF tests on yt and xt .
Critical value tρb: -3.4 (if extra term βt: -3.8)
→ Take difference until non-stationarity is rejected.
Step 2: Estimate ADL model for stationary series by OLS. Error Correction Model (ECM): if xt and yt cointegrated, estimate
∆yt = α + β1 (yt−1 − bxt−1 ) + β2 ∆yt−1 + β3 ∆xt−1 + εt
One exception: if xt and yt are cointegrated. (or more lags for ∆yt and ∆xt )
erasmus.png erasmus.png
Lecture 6.3, Slide 11 of 13, Erasmus School of Economics Lecture 6.3, Slide 12 of 13, Erasmus School of Economics
TRAINING EXERCISE 6.3

Train yourself by making the training exercise (see the website).

After making this exercise, check your answers by studying the

webcast solution (also available on the website).

erasmus.png
Lecture 6.3, Slide 13 of 13, Erasmus School of Economics