Dr.

Rudra Pradhan

VINOD GUPTA SCHOOL OF MANAGEMENT

INDIAN INSTITUTE OF TECHNOLOGY, KHARAGPUR

Subject: Econometric Modelling

Assignments: 6-8

Name: Wagh Abhishek Satish

Roll Number: 18BM60150
=======================================================================
Q 1. If Yt = a + d Yt-1 + U1, find out ∆∆∆Yt
Ans 1: ∆∆∆Yt = d ∆∆∆Yt-1 + ∆∆∆Ut

Q 2. If the model is of ARIMA (p, q, r),

Define p
Define q
Define r
Is there any suitable strategy to find optimum value of p, q and r; if so, explain it.
Ans 2:
p: The lag order, i.e. the number of lag observations included in the model
q: The degree of differencing, i.e. the number of times that the raw observations are differenced
r: The order of moving average, i.e. the size of the moving average window

Strategy to find out the optimal value of p, q and r.

For finding “q”
• ADF test or any other test of stationarity to check the order of integration for which the given time series
becomes stationary.
• The order of differencing for which the time series becomes stationary is the value of “q”

For finding “p” and “r”

• Plot the Autocorrelation Plot (ACF) and Partial Autocorrelation Plot (PACF) of the time series which is
differentiated “q” times.
• If the PACF of the differenced series displays a sharp cutoff and/or the lag-1 autocorrelation is positive--
i.e., if the series appears slightly "under-differenced"--then consider adding an AR term to the model. The
lag at which the PACF cuts off is the indicated number of AR terms.
• If the ACF of the differenced series displays a sharp cutoff and/or the lag-1 autocorrelation is negative--
i.e., if the series appears slightly "over-differenced"--then consider adding an MA term to the model. The
lag at which the ACF cuts off is the indicated number of MA terms.

Q 3. The estimated model is Yt = β – W1 et-1 – W2 et-2. If β = 75.4; W1 = 0.57 and W2 = -0.36, find
out Y76.
(a) 80.6 (b) 90.6
(c) 70.6 (d) None
Ans 3: Cannot be answered since the actual values for Yt are unknown, which in turn are required to calculate
error terms.
Q 4. Which is not a time series model?
(a) AR (b) ARMA
(c) ARDL (d) LOGIT
Ans 4: (d) LOGIT

Q 5. Which is not used for Model specification?

(a) AIC (b) SBC
(c) R2 (d) none
2
Ans 5: (c) R

Q 6. Which is not used to check the stationarity check?

(a) PP test (b) DE test
(c) ADF test (d) None
Ans 6: (b) DE test

Q 7. If Yt = a + b Yt-1 + ut, write its second difference model and its financial implications.
Ans 7: ∆∆Yt = b ∆∆Yt-1 + ∆∆ut
Financial implications cannot be given as Yt is not defined.

Q 8. What is not used for cointegration check?

(a) EG test (b) JJ test
(c) R2 = 0 (d) None
Ans 8: (b) R2 = 0

Q 9. VECM is used to detect

(a) Multicollinearity (b) Serial Correlation
(c) None (d) Identification
Ans 9: (b) Serial Correlation

Q 10. Which is not volatility modelling group?

(a) ARCH (b) GARCH
(c) EGARCH (d) None
Ans 10: (d) None

Q 11. Which model is used for cointegration check?

(a) ARCH (b) IGARCH
(c) EGARCH (d) None
Ans 11: (d) None

Q 12. Which is not a part of Out-of-sample prediction test

(a) MAD (b) RMSE
(c) MAPE (d) None
Ans 12: (a) MAD

Q 13. ARDL
(b) Auto regressive dummy lag modelling (b) Average regressive Distributive Model
(c) None (d) Auto regressive Distributive Lag Model
Ans 13: (d) Auto regressive Distributive Lag Model
Q 14. Find the rank of this matrix [12 15]
(a) 1 (b) 0
(c) 2 (d) None
Ans 14: (a) 1

Q 15. LM Test is meant to find out

(a) ARCH effect (b) unit root
(c) Cointegration (d) None
Ans 15: (a) ARCH effect

Q 16. Follow the below regression results:

Model 1 lg Mt = -10.26 + 1.60 lg Gt
[-12.0] [26.0]

R2 = 0.95 and d = 0.33

Model 2 ∆lg Mt = 0.001 + 0.58 ∆lg Gt

[2.50] [1.90]

R2 = 0.09 and d = 1.74

Model 3 ∆ut = -0.195 ut-1

[-2.253]

R2 = 0.12 and d = 1.50

Now answer the followings:

A. Interpret the regression model 1?

Ans 16.A: According to Granger and Newbold, an R2 > d is a good rule of thumb to suspect that the
estimated regression is spurious. Here as well, R2 (0.95) > d (0.33), it is a sign of spurious
regression. Also, such a small value of Durbin-Watson (d) statistic shows there is first order
autocorrelation. Additional data about order of integration of Lg Mt and Lg Gt should be provided.
This information is crucial because both the series should be non-stationary at level and stationary at
first difference, to talk about spurious regression.

B. Interpret the regression model 1?

Ans 16.B:
Calculate Statistics: d = 0.33
Tabulated Statistics: Assumption: n (number of observations) = 20
Given: k’ (number of explanatory variables including constant) = 3
dL = 0.998, dU = 1.676
Probability Level: α = 0.05
 C. Do you suspect any spurious regression; why?
Ans 16.C: In Model 1, where both variables are level (not difference), R2 is abnormally high. Model
2 is relation between 1st difference of both series. The R2 value tends to zero and d is within range
(1.5 to 2.5). According to Granger and Newbold, an R2 > d is a good rule of thumb to suspect that
the estimated regression is spurious. In Model 1, R2 (0.95) > d (0.33), we only suspect (not sure)
spurious regression.

D. Interpret the model 3 results, can it be used to detect spurious issue?

Ans 16.D: Yes, it can. Non-stationary residuals is one of the signs of spurious regression. In Model
3, the t-value of coefficient of ut-1 -2.253. The critical DF values for such model (w/o drift and trend)
are as follows:

From the critical values tabulated, it is clear that the residuals series is stationary at α = 0.1 and α =
0.05. But the it is non-stationary at α = 0.01. Even if we consider α = 0.05, the conclusion could be
that the residuals are stationary and hence no spurious regression. In answers A, B and C above, we
have used a rule of thumb and have used an assumption of degree of freedom. Since residuals are
stationary, we conclude that it is not a spurious regression.
Q 17.

Yt : 2.8 3.4 3.0 3.5 3.6 3.0 2.7 3.7

Xt : 1.8 2.4 4.0 2.5 1.6 4.0 5.7 6.7

Time : 1 2 3 4 5 6 7 8

Now answer the followings:

A. Find out DF statistics of X and Y:
Ans 17.A: Using MS Excel, we regress ΔXt on Xt-1. DF statistics is nothing but t-stat of coefficient
of Xt-1 = -0.429526883722013
Similarly, we regress ΔYt on Yt-1. DF statistics is nothing but t-stat of coefficient
of Yt-1 = -3.03455911785909

B. Comment the stationary issue?

Ans 17.B: Using EViews® 10+ Student Version Lite, we carry out ADF test on Xt and Yt separately.
Results are as follows:
Null Hypothesis: X has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 1 (Automatic - based on SIC, maxlag=1)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -1.621575 0.6717

Test critical values: 1% level -7.006336
5% level -4.773194
10% level -3.877714

*MacKinnon (1996) one-sided p-values.

Warning: Probabilities and critical values calculated for 20 observations
and may not be accurate for a sample size of 6
Null Hypothesis: D(X) has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 1 (Automatic - based on SIC, maxlag=1)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -9.492802 0.0065

Test critical values: 1% level -8.235570
5% level -5.338346
10% level -4.187634

*MacKinnon (1996) one-sided p-values.

Warning: Probabilities and critical values calculated for 20 observations
and may not be accurate for a sample size of 5
The above results show that Xt is I(1).
Similarly, for Yt:
Null Hypothesis: Y has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 1 (Automatic - based on SIC, maxlag=1)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -1.922559 0.5228

Test critical values: 1% level -7.006336
5% level -4.773194
10% level -3.877714

*MacKinnon (1996) one-sided p-values.

Warning: Probabilities and critical values calculated for 20 observations
and may not be accurate for a sample size of 6
Null Hypothesis: D(Y) has a unit root
Exogenous: Constant, Linear Trend
Lag Length: 1 (Automatic - based on SIC, maxlag=1)

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -1.697683 0.6507

Test critical values: 1% level -8.235570
5% level -5.338346
10% level -4.187634

*MacKinnon (1996) one-sided p-values.

Warning: Probabilities and critical values calculated for 20 observations
and may not be accurate for a sample size of 5
ADF on ΔΔYt cannot be carried out due to insufficient observations. Hence, we can conclude that
order of integration of Yt is greater than 1.

C. Are X and Y cointegrated?

Ans 17.C: If the residuals of regression of Yt over Xt are stationary, it can be concluded that Yt and
Xt are cointegrated. We calculated the regression model as:
Yt = 3.258215962 - 0.0127431254191817 Xt + ut.
From this, we generate series ut and run unit root test on it in EViews. The result is that at α = 0.05,
the series ut is not stationary. Hence, we can conclude that X & Y are not cointegrated.

D. Find out ECM value

Ans 17.D:
Since only Xt is I(1), it could be inappropriate to calculate ECM using Yt, whose order of integration
is unknown. Still, if we go ahead and find ECM, we get following in EViews® 10+ Student Version
Lite:
Dependent Variable: D(Y)
Method: Least Squares
Date: 03/28/20 Time: 15:15
Sample (adjusted): 2 8
Included observations: 7 after adjustments

Variable Coefficient Std. Error t-Statistic Prob.

C 0.153060 0.142448 1.074497 0.3431

D(X) -0.171452 0.092810 -1.847348 0.1384
UT(-1) -1.268477 0.385476 -3.290681 0.0302

R-squared 0.806487 Mean dependent var 0.128571

Adjusted R-squared 0.709731 S.D. dependent var 0.593617
S.E. of regression 0.319821 Akaike info criterion 0.855414
Sum squared resid 0.409141 Schwarz criterion 0.832233
Log likelihood 0.006051 Hannan-Quinn criter. 0.568897
F-statistic 8.335245 Durbin-Watson stat 2.669912
Prob(F-statistic) 0.037447

Hence, ECM is ΔY = 0.15306 – 0.171452 ΔXt – 1.268477 ut-1 + v

E. Comment on GC test?
Ans 17.E: The pre-requisite of Granger Causality test is that the two series must be stationary. Due
to insufficient readings in Yt, it cannot be made stationary. Still, if we carry out Granger causality
test between ΔYt,[assuming Yt is I(1)] and ΔXt [since Xt is I(1)] in EViews® 10+ Student Version
Lite, results are:

Pairwise Granger Causality Tests

Date: 03/28/20 Time: 16:19
Sample: 1 8
Lags: 1

Null Hypothesis: Obs F-Statistic Prob.

D(X) does not Granger Cause D(Y) 6 0.05017 0.8372

D(Y) does not Granger Cause D(X) 0.04202 0.8507

Above result shows that there is no causality between ΔYt and ΔXt. More than one lag cannot be
included due to insufficient sample size.
Q 18. Follow the below regression results:

Model 1 ∆Mt = 0.03Mt-1

[12.7] R2 = 0.95; RSS = 207 and d = 0.37

Model 2 ∆Mt = 1.87 + 0.02Mt-1

[3.3] [3.9] R2 = 0.25; RSS = 167 and d = 0.45

Model 3 ∆Mt = 1.16 + 0.53t - 0.11Mt-1

[2.4] [4.8] [-4.0] R2 = 0.51; RSS = 110 and d = 0.61

Comment the regression results?

Ans 18:
Model 1: δ > 0 ⇒ ρ > 1 ⇒ series is explosive. Hence, this model without drift and deterministic trend
should be eliminated.

Model 2: δ > 0 ⇒ ρ > 1 ⇒ series is explosive. Hence, this model without deterministic trend should
be eliminated.

Model 3: δ = -0.11. It is less than zero, but is it statistically significantly less than zero? Here, τ = -
4.0 & |τ| = 4.0. From Table D.7 of D N Gujarati textbook, for α = 0.05, δ is significant for sample
size ≥ 25. Hence, this model shows that Mt is stationary at level i.e. I(0).
Q 19.

X: 2 5 15 25 35 45 55 50 45 60 85

Y: 8 10 20 30 50 70 80 65 55 85 96

Given the above information, answer the following:

1. Examine unit root test of these two series
2. What is there order of integration where they attain stationarity?
3. Suggest suitable lag length for this estimation
4. Design ACF and PACF for both the series
5. plot these two series and check the volatility issue
6. Fit appropriate ARIMA model

Ans 19:
1. ADF test to check for unit root. EViews® 10+ Student Version Lite is used.
For series X:
t-Statistic Prob.*

Augmented Dickey-Fuller test statistic 0.183667 0.9551

Test critical values: 1% level -4.297073
5% level -3.212696
10% level -2.747676

As p-value is 0.95, we fail to reject H0 and conclude that the series has a unit root and is hence non-
stationary at level
For series Y:
t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -0.671397 0.8104

Test critical values: 1% level -4.297073
5% level -3.212696
10% level -2.747676

As p-value is 0.81, we fail to reject H0 and conclude that the series has a unit root and is hence non-
stationary at level

2. For series X, at first difference we get following results for ADF test:

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -3.510504 0.0390

Test critical values: 1% level -4.582648
5% level -3.320969
10% level -2.801384

We see that we can reject the null hypothesis at 5% significance level and conclude that the series
becomes stationary after first order of differencing
For series Y, at first difference we get following results for ADF test:

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -2.907739 0.0865

Test critical values: 1% level -4.582648
5% level -3.320969
10% level -2.801384

At second difference, we get following results for ADF test:

t-Statistic Prob.*

Augmented Dickey-Fuller test statistic -3.565003 0.0412

Test critical values: 1% level -4.803492
5% level -3.403313
10% level -2.841819

Hence, we can reject H0 at second difference and conclude that the series becomes stationary at
second order of differencing, i.e. I(2).

3. For series X:

VAR Lag Order Selection Criteria

Endogenous variables: X
Exogenous variables: C
Date: 03/26/20 Time: 14:26
Sample: 1 11
Included observations: 7

Lag LogL LR FPE AIC SC HQ

0 -28.80051 NA 292.5170 8.514432 8.506705 8.418926

1 -25.97850 4.031449 176.3265 7.993857 7.978402 7.802845
2 -25.46903 0.582253 211.7229 8.134007 8.110826 7.847490
3 -18.36166 6.092028* 40.75565* 6.389046* 6.358137* 6.007023*
4 -18.35309 0.004898 66.52793 6.672311 6.633675 6.194782

We see that 4 of 5 criteria point out lag upto 3 as optimum. Hence suitable lag length for X is 3.
For series Y:

VAR Lag Order Selection Criteria

Endogenous variables: Y
Exogenous variables: C
Date: 03/26/20 Time: 14:29
Sample: 1 11
Included observations: 6

Lag LogL LR FPE AIC SC HQ

0 -24.12244 NA* 254.5278 8.374148 8.339441 8.235214

1 -23.90483 0.290158 338.1689 8.634942 8.565528 8.357074
2 -23.07065 0.834177 384.1186 8.690216 8.586096 8.273414
3 -18.83511 2.823690 156.0119* 7.611705 7.472878 7.055969
4 -18.83071 0.001467 342.7229 7.943571 7.770037 7.248901
5 153.8996 0.000000 NA -49.29986* -49.50810* -50.13346*

From the above table lag 5 seems to be optimal lag length for series Y.

4. ACF and PACF for series X:

ACF and PACF for series Y:

5. Graph of X:
X
90

80

70

60

50

40

30

20

10

0
1 2 3 4 5 6 7 8 9 10 11
Volatility of X: 5.012

Graph of Y:
Y
100

80

60

40

20

0
1 2 3 4 5 6 7 8 9 10 11
Volatility of Y: 5.55

6. We know that the order of differencing for X is 2 and that for Y is 3 to make them stationary.
Hence we plot the ACF and PACF for the differenced X and Y.
 Fig. ACF and PACF for X with first order differencing
We see that the ARIMA(2,1,0) model is optimal for X.
Model for X:

Dependent Variable: D(X)

Method: ARMA Maximum Likelihood (OPG - BHHH)
Date: 03/26/20 Time: 16:38
Sample: 2 11
Included observations: 10
Convergence achieved after 14 iterations
Coefficient covariance computed using outer product of gradients

Variable Coefficient Std. Error t-Statistic Prob.

C 7.334010 2.281102 3.215117 0.0182

AR(1) 0.545085 0.331990 1.641871 0.1517
AR(2) -0.832723 0.273638 -3.043152 0.0227
SIGMASQ 22.37434 18.84718 1.187145 0.2800

R-squared 0.689288 Mean dependent var 8.300000

Adjusted R-squared 0.533933 S.D. dependent var 8.944893
S.E. of regression 6.106600 Akaike info criterion 6.991514
Sum squared resid 223.7434 Schwarz criterion 7.112548
Log likelihood -30.95757 Hannan-Quinn criter. 6.858740
F-statistic 4.436838 Durbin-Watson stat 1.895658
Prob(F-statistic) 0.057423
 Fig. ACF and PACF for Y with second order differencing
We see that the ARIMA(5,1,0) model is optimal for Y.
Model for Y:

Dependent Variable: D(Y,2)

Method: ARMA Maximum Likelihood (OPG - BHHH)
Date: 03/26/20 Time: 16:39
Sample: 3 11
Included observations: 9
Convergence achieved after 35 iterations
Coefficient covariance computed using outer product of gradients

Variable Coefficient Std. Error t-Statistic Prob.

C -1.897450 1.085368 -1.748209 0.2225

AR(1) -0.907635 0.412866 -2.198376 0.1590
AR(2) -1.494522 0.747061 -2.000535 0.1834
AR(3) -1.322625 0.606810 -2.179636 0.1611
AR(4) -0.884808 0.547321 -1.616616 0.2473
AR(5) -0.661358 2.060291 -0.321002 0.7786
SIGMASQ 69.18370 180.2281 0.383867 0.7380

R-squared 0.782745 Mean dependent var 1.000000

Adjusted R-squared 0.130979 S.D. dependent var 18.92749
S.E. of regression 17.64445 Akaike info criterion 9.290963
Sum squared resid 622.6533 Schwarz criterion 9.444359
Log likelihood -34.80933 Hannan-Quinn criter. 8.959933
F-statistic 1.200961 Durbin-Watson stat 2.037051
Prob(F-statistic) 0.520420
Q 20.
Scores of a series (Y): 2.8 3.4 3.0 3.5 3.6 3.0 2.7 3.7 4.5 4.2

Time: 1 2 3 4 5 6 7 8 9 10

Critical value: assume 2. 0 for simple t- and F-stat and -2.0 for DF and ADF stat

Let the estimated model is like this:

P = 34.57 + 0.792 t
[6.358] [0.567]

A. Using the above series fit the AR model of order 2, 2.

B. Using the above series fit the MA model of order 2, 2.
C. Find out ACF at P = 2
D. Can you fit an ARIMA model here? If not, what is the difficulty?

Ans 20:
Ans 20.A: AR(2) model for the given series:

Dependent Variable: Y
Method: ARMA Maximum Likelihood (OPG - BHHH)
Date: 03/26/20 Time: 17:23
Sample: 1 10
Included observations: 10
Convergence achieved after 12 iterations
Coefficient covariance computed using outer product of gradients

Variable Coefficient Std. Error t-Statistic Prob.

C 3.095475 1.327200 2.332336 0.0585

AR(1) 0.146529 1.624963 0.090174 0.9311
AR(2) -0.240517 2.096533 -0.114721 0.9124
SIGMASQ 1.271177 1.359509 0.935027 0.3859

R-squared 0.084430 Mean dependent var 3.140000

Adjusted R-squared -0.373355 S.D. dependent var 1.242041
S.E. of regression 1.455551 Akaike info criterion 3.891144
Sum squared resid 12.71177 Schwarz criterion 4.012178
Log likelihood -15.45572 Hannan-Quinn criter. 3.758369
F-statistic 0.184431 Durbin-Watson stat 2.025602
Prob(F-statistic) 0.903216

Inverted AR Roots .07+.48i .07-.48i

Ans 20.B: MA(2) model for the given series:
Dependent Variable: Y
Method: ARMA Maximum Likelihood (OPG - BHHH)
Date: 03/26/20 Time: 17:25
Sample: 1 10
Included observations: 10
Convergence achieved after 78 iterations
Coefficient covariance computed using outer product of gradients

Variable Coefficient Std. Error t-Statistic Prob.

C 3.003014 0.272205 11.03218 0.0000

MA(1) -0.347382 23869.80 -1.46E-05 1.0000
MA(2) -0.652617 65295.23 -9.99E-06 1.0000
SIGMASQ 0.899406 7066.590 0.000127 0.9999

R-squared 0.352199 Mean dependent var 3.140000

Adjusted R-squared 0.028299 S.D. dependent var 1.242041
S.E. of regression 1.224341 Akaike info criterion 3.733622
Sum squared resid 8.994064 Schwarz criterion 3.854656
Log likelihood -14.66811 Hannan-Quinn criter. 3.600848
F-statistic 1.087370 Durbin-Watson stat 1.833182
Prob(F-statistic) 0.423417

Inverted MA Roots 1.00 -.65

Ans 20.C: ACF and PACF at level:

ACF and PACF at second order differencing:

D. Yes, an ARIMA model can be fit on the series Y.

Q 21.
Consider the time series model:

Yt = λ +βYt-1 + ut

A. Calculate the unconditional Mean of Yt

B. Calculate the unconditional variance of Yt
C. Derive the ACF for this process.
D. If λ = 150 and β = -0.5 and if the current observation is Y100 = 85, would you expect the
next observation to be above or below the mean?

Ans 21:
Ans 21.A:
The unconditional mean is given by the expected value of Yt
E(Yt) = λ + β E(Yt−1) + E(ut)

But E(ut)=0 as mean of error term is zero

⇒ E(Yt) = λ + βE(Yt−1)

Now Yt-1 = λ +β Yt-2 + ut-1. Substituting this in above equation:

⇒ E(Yt) = λ + β E(λ +β Yt-2 + ut-1)
⇒ E(Yt) = λ + β[λ +βE (Yt-2)]
⇒ E(Yt) = λ+ β λ+ β2 E(Yt-2)

Now Yt-2 = λ +β Yt-3 + ut-2. Substituting this in above equation:

⇒ E(Yt) = λ+ β λ+ β2 E (λ +βYt-3 + ut-2)
⇒ E(Yt) = λ+ β λ+ β2 λ + β3 E(Yt-3)

We can notice a pattern here, going on making such substitutions would give
⇒ E(Yt) = λ (1+ β + β2 + … β(n-1)) + βn E(Yt-n)

So long as the series is stationary i.e. | β |<1, β∞=0, Therefore, taking limits as n tends to infinity,
E(Yt-n) tends to 0. Substituting this in above equation,
⇒ E(Yt) = λ(1+ β + β2 + … β(n-1))

Sum of the series 1+x+x2+x3+…… is given by (1/(1-x)). Hence,

⇒ E(Yt) = λ/(1- β) which is the unconditional mean of Yt

Ans 21.B:
Var(yt) = β2Var(yt−1) + Var(ut)
⇒ Var(yt) = β2Var(y) + σ2
⇒ Var(y) = σ2/(1 − β2 )
Ans 21.C:
Turning now to the calculation of the autocorrelation function, the autocovariances must first be
calculated. This is achieved by following similar algebraic manipulations as for the variance above,
starting with the definition of the autocovariances for a random variable.

The autocovariances for lags 1, 2, 3, . . . , s, will be denoted by γ1, γ2, γ3,..., γs , as previously.

γ1 = cov (yt, yt−1) = E[yt − E(yt)][yt−1 − E(yt−1)]

Since μ has been set to zero,

E(yt) = 0 and E(yt−1) = 0, so γ1 = E[yt yt−1]

under the result above that E(yt) = E(yt−1) = 0. Thus,

γ1 = E [(ut + φ1ut−1 + φ2 1 ut−2 +···) (ut−1 + φ1ut−2 + φ21 ut−3 +···)]
γ1 = E [ φ1u2t−1 + φ31 u2t−2 +···+ cross – products]

Again, the cross-products can be ignored so that

γ1 = φ1σ2 + φ31σ2 + φ51σ2 +···
γ1 = φ1σ2 (1 + φ21 + φ41 +···)
γ1 = φ1σ2 / (1 − φ21)

For the second autocovariance,

γ2 = cov (yt, yt−2) = E[yt − E(yt)][yt−2 − E(yt−2)]

Using the same rules as applied above for the lag 1 covariance
γ2 = E[yt yt−2]
γ2 = E[( ut + φ1ut−1 + φ21 ut−2 +··· ) (ut−2 + φ1ut−3 + φ21 ut−4 +··· )]
γ2 = E (φ21 u2t−2 + φ41 u2t−3 +···+cross-products)
γ2 = φ21σ2 + φ41σ2 +···
γ2 = φ21σ2 (1 + φ21 + φ41 +···)
γ2 = φ21σ2 /(1 − φ21)

By now it should be possible to see a pattern emerging. If these steps were repeated for γ3, the
following expression would be obtained
γ3 = φ31σ2/ (1 − φ21)

Univariate time series modelling and forecasting and for any lag s, the autocovariance would be
given by
γs = φs1σ2 / (1 − φ21)

The acf can now be obtained by dividing the covariances by the variance, so that
τ0 = γ0 / γ0 = 1
τ1 = γ1 /γ0 = (φ1σ2 / (1 − φ21)) / (σ2 / (1 − φ21)) = φ1
τ2 = γ2 /γ0 = (φ21σ2 / (1 − φ21))/( σ2 /(1 – φ21)) = φ21
τ3 = φ31
The autocorrelation at lag s is given by
τs = φs1 which means that
corr (yt, yt−s) = φs1.

Ans 21.D:
As derived in section A of the answer, the unconditional mean is given by:
E(Yt) = λ/(1- β) = 150/(1+0.5) = 100

The next observation is given by Y101 = 150 - 0.5 * 85 = 107.5

Hence the next observation is above the mean.

==XX==
The End