First evaluation step: Check for stationarity

Take difference of time series until stationarity.

MOOC Econometrics Test equation: Augmented Dickey-Fuller
Lecture 6.4 on Time Series: ∆yt = α + βt + ρyt−1 + γ1 ∆yt−1 + . . . + γL ∆yt−L + εt
Evaluation and Illustration Critical value tρb: −2.9 if β = 0, −3.5 if β 6= 0
Dick van Dijk, Philip Hans Franses, Christiaan Heij
For stationary data:
Pp
→ OLS in AR: yt = α + j=1 βj yt−j + εt
Pp
with trend: yt = α + γt + j=1 βj yt−j + εt

→ OLS in ADL: yt = α + pj=1 βj yt−j + rj=1 γj xt−j + εt

P P

with trend: yt = α + δt + pj=1 βj yt−j + rj=1 γj xt−j + εt

P P

t- and F -tests as usual.

Lecture 6.4, Slide 2 of 12, Erasmus School of Economics

Check for cointegration Diagnostic tests

If xt and yt are both non-stationary: check for cointegration. Choice of lag lengths: BIC (see Lecture 3).

Test method: Engle-Granger two-step method Stability check: Chow tests (see Lecture 3).

→ OLS in yt = α + βxt + εt → b and OLS residuals et

Normal residuals: Jarque-Bera (see Lecture 3), critical value: 6.0.
→ OLS in ∆et = α + βt + ρet−1 + γ1 ∆et−1 + . . . + γL ∆et−L + ωt
Critical value tρb: −3.4 if β = 0, −3.8 if β 6= 0 Out-of-sample forecasting: Lecture 6.5.

If xt and yt are cointegrated, estimate ECM: Model should in particular capture autocorrelation in time series.
∆yt = α+βt+γ0 (yt−1 −bxt−1 )+ pj=1 γy ,j ∆yt−j + rj=1 γx,j ∆xt−j +εt
P P → Test if model residuals are uncorrelated: white noise.

(or with β = 0) Two tests: ACF and Breusch-Godfrey.

√
t- and F -tests as usual. ACF rule-of-thumb: significant if |ACF| > 2/ n.
Lecture 6.4, Slide 3 of 12, Erasmus School of Economics Lecture 6.4, Slide 4 of 12, Erasmus School of Economics
Test question Test on serial correlation: Breusch-Godfrey
Test
Let yt be white noise with variance σ 2 . Show that OLS estimator b in Step 1: Estimate model and get residuals et .
yt = α + βyt−1 + εt gives the first-order autocorrelation of yt . Further
√ √
show that (−2/ n, 2/ n) is approximate 95% confidence interval for β. Step 2: Regress et on all variables of model and r lags of et .
Hint: Use results of Lecture 1.
Answer: Step 3: BG = nR 2 of Step 2, and BG ≈ χ2 (r ) if et white noise.

yt = α = βxt + εt where xt = yt−1 , t = 2, . . . , n, so

Example: Model yt = α + βyt−1 + γxt−1 + εt
b = nt=2 (yt − y )(yt−1 − y )/ nt=2 (yt−1 − y )2
P P

→ Step 1: OLS residuals et = yt − a − byt−1 − cxt−1 .

Pn
var(b) = σ2/ t=2 (yt−1 − where y )2 , → Step 2: OLS in et = α + βyt−1 + γxt−1 + δ1 et−1 + δ2 et−2 + ωt
Pn 2
P n 2 2
t=2 (yt−1 − y ) = (n − 1) t=2 (yt−1 − y ) /(n − 1) ≈ (n − 1)σ , → Step 3: BG = nR 2 ≈ χ2 (2) if et white noise.
var(b) ≈ σ 2 /((n − 1)σ 2 ) = 1/(n − 1) ≈ 1/n
→ Conclusion: Model not correctly specified if BG > 6.0.
√ → Should then adjust model, e.g. more lags of yt and xt .
If n large then b ≈ 0 and SE(b) ≈ 1/ n
√ √
b − 2SE(b) < β < b − 2SE(b) → Lecture
−2/ n < β < 2/ n
6.4, Slide 5 of 12, Erasmus School of Economics Lecture 6.4, Slide 6 of 12, Erasmus School of Economics

Illustration: Revenue Passenger Kilometers (RPK) Tests on stationarity

3.4 Let yt denote log(RPK), either X1t or X2t : trend
X1 = log(RPK1)
3.2
X2 = log(RPK2)
ADF: ∆yt = α + βt + ρyt−1 + γ∆yt−1 + εt
3.0

2.8 t-value of ρb: t = −2.8 for X1 , t = −1.2 for X2

2.6

2.4
Let yt denote either ∆X1t or ∆X2t : no trend
2.2
1975 1980 1985 1990 1995 2000 2005 2010 2015

YEAR ADF: ∆yt = α + ρyt−1 + γ∆yt−1 + εt

DX1 DX2
t-value of ρb: t = −3.3 for X1 , t = −3.7 for X2
.04 .04

Test
.02 .02

.00 .00
What conclusions do you draw from these outcomes?
-.02 -.02
Answer:
1970 1980 1990 2000 2010 2020 1970 1980 1990 2000 2010 2020

YEAR YEA R
As t > −3.5, X1 and X2 not stationary.
Graphs suggest: X1 and X2 non-stationary, ∆X1 and ∆X2 stationary.
As t < −2.9, ∆X1 and ∆X2 are both stationary.
Lecture 6.4, Slide 7 of 12, Erasmus School of Economics Lecture 6.4, Slide 8 of 12, Erasmus School of Economics
Granger causality tests Engle-Granger test and ECM
Step 1: OLS: X2t = 0.01 + 0.92X1t + et .

ADL for ∆X1t ADL for ∆X2t

Step 2: ADF: ∆et = 0.00 − 0.50et−1 + 0.30∆et−1 + rest
Coef. t-Stat. p-value Coef. t-Stat. p-value
→ t-value of coefficient et−1 : t = −3.5 < −3.4
Constant 0.01 1.85 0.07 0.01 2.86 0.01 → et stationary → X1t and X2t cointegrated.
∆X1,t−1 0.87 4.96 0.00 0.18 1.29 0.21
∆X1,t−2 -0.42 -2.02 0.05 0.61 3.68 0.00 ECM (after removing insignificant coefficients):
∆X2,t−1 0.35 1.74 0.09 -0.29 -1.81 0.08
∆X1t = 0.00 + 1.02∆X1,t−1 + 0.46(X2,t−1 − 0.92X1,t−1 ) + e1t
∆X2,t−2 -0.19 -1.27 0.21 -0.13 -1.05 0.30
∆X2t = 0.02 − 0.45(X2,t−1 − 0.92X1,t−1 ) + e2t

Company 1 Granger causal for company 2, not other way round. If Dt−1 = X2,t−1 − 0.92X1,t−1 is positive, then
→ See t-tests (confirmed by F -tests on two coefficients jointly). 0.46 > 0 → X1t ↑ → Dt = X2t − 0.92X1t ↓
−0.45 < 0 → X2t ↓ → Dt = X2t − 0.92X1t ↓

Lecture 6.4, Slide 9 of 12, Erasmus School of Economics Error correction mechanism acts on both variables.
Lecture 6.4, Slide 10 of 12, Erasmus School of Economics

ECM: Check for serial correlation and normality TRAINING EXERCISE 6.4
ECM models for log(RPK) of airline companies 1 and 2 (n = 39):
∆X1t = 0.00 + 1.02∆X1,t−1 + 0.46(X2,t−1 − 0.92X1,t−1 ) + e1t
∆X2t = 0.02 − 0.45(X2,t−1 − 0.92X1,t−1 ) + e2t

Jarque-Bera test: JB1 = 0.4 < 6, JB2 = 1.8 < 6. Train yourself by making the training exercise (see the website).
Breusch-Godfrey test (1 lag): BG1 = 0.3 < 3.9, BG2 = 1.2 < 3.9.
√ √ After making this exercise, check your answers by studying the
ACF: 2/ n = 2/ 39 = 0.32.
.5
ACF e1
webcast solution (also available on the website).
.4 ACF e2
.3

.2

.1

.0

-.1

-.2

-.3
0 2 4 6 8 10

LAG

Lecture 6.4, Slide 11 of 12, Erasmus School of Economics Lecture 6.4, Slide 12 of 12, Erasmus School of Economics