lOMoARcPSD|3191521

Tutorial work, Time series analysis, Assignment 3, Solutions

Tijdreeksanalyse (Erasmus Universiteit Rotterdam)

StuDocu is not sponsored or endorsed by any college or university

Downloaded by Marita Nadyah ([email protected])
 lOMoARcPSD|3191521

BSc Econometrics
FEB23001(X)-13 Time Series Analysis
Assignment 3 - Solutions

ASSIGNMENT 3.1
3.1.a) [1 point] The time series of quarterly growth rates (in percent) of consumption
is shown below (for the complete sample period), as well as the histogram (plus a
normal density with the same mean and variance) for the sample period 1970Q1-
1999Q4. Skewness of the growth rates over this period is equal to −1.04, while
kurtosis is equal to 5.74. Both are very different from the normal values of 0 and
3, such that the null hypothesis of normality is convincingly rejected when applying
the Jarque-Bera test: The test statistic takes a value of 59.0, with p-value 0.000.
3

-1

-2

-3
1970 1975 1980 1985 1990 1995 2000 2005 2010

Figure 1: Quarterly growth rates of US consumption, 1968Q4-2013Q4.

Shaded areas are NBER-defined recessions
.9

.8 Histogram
Normal
.7

.6
Density

.5

.4

.3

.2

.1

.0
-3 -2 -1 0 1 2 3 4 5

Figure 2: Histogram of quarterly consumption growth rates, 1970Q1-1999Q4.

The first 12 (partial) autocorrelations are shown below, with an asterisk indicating
statistical significance at the 5% level. We find significant autocorrelations at lages 1,
2, 3, and 8, and significant partial autocorrelations at lags 1 and 3 (but not at lag 2).

Downloaded by Marita Nadyah ([email protected])

 lOMoARcPSD|3191521

Original time 1974Q4 and 1980Q2

series set to 0
k AC PAC AC PAC
∗
1 0.259 0.259∗ 0.281 ∗
0.281∗
2 0.229∗ 0.173 0.297∗ 0.237∗
3 0.265∗ 0.189∗ 0.295 ∗
0.189∗
4 0.064 −0.073 −0.035 −0.234∗
5 −0.004 −0.094 −0.025 −0.115
6 0.013 −0.020 −0.090 −0.073
7 −0.091 −0.078 −0.112 0.034
8 −0.198∗ −0.162 −0.159 −0.089
9 −0.092 0.005 −0.105 −0.006
10 −0.106 0.008 −0.035 0.056
11 −0.077 0.043 −0.071 −0.009
12 −0.130 −0.111 −0.081 −0.119

3.1.b) [2 points] We estimate an AR(4) model φ4 (L)(yt − µ) = εt , where yt denotes

the quarterly consumption growth rate, and φ4 (L) = 1 − φ1 L − φ2 L2 − φ3 L3 − φ4 L4 .
Using the observations for t =1970Q1-1999Q4, we find the following least squares
estimates, with standard errors in parentheses: µ̂ = 0.860(0.119), φ̂1 = 0.196(0.093),
φ̂2 = 0.144(0.093), φ̂3 = 0.208(0.093), and φ̂4 = −0.072(0.093). Note that the esti-
mated coefficients for the second and fourth lag are not significantly different from
zero even at the 10% level. The R2 of the model is equal to 0.135.

The residuals from the AR(4) model have skewness equal to −0.89 and kurtosis
equal to 5.50. Normality is clearly rejected as the Jarque-Bera statistics is equal
to 47.0, with a p-value of 0.000. The (partial) autocorrelations of the residuals are
shown below; for both, we find a significant value at k = 8. No significant (partial)
autocorrelations appear for the squared residuals, such that heteroskedasticity does
not seem to be an issue.

AR(4) residuals AR(4)-AO residuals AR(4)-IO residuals

k AC PAC AC PAC AC PAC
1 −0.006 −0.006 −0.031 −0.031 0.065 0.065
2 0.015 0.015 −0.012 −0.013 0.136 0.133
3 0.011 0.011 0.063 0.062 0.050 0.034
4 0.030 0.030 0.020 0.024 −0.095 −0.120
5 −0.033 −0.033 −0.092 −0.090 −0.088 −0.091
6 0.018 0.017 −0.070 −0.081 −0.138 −0.106
7 −0.039 −0.038 0.046 0.038 −0.047 −0.001
8 −0.190∗ −0.192∗ −0.101 −0.090 −0.121 −0.091
9 −0.011 −0.012 −0.007 0.001 0.000 0.011
10 −0.037 −0.034 0.062 0.052 0.067 0.071
11 0.049 0.058 −0.036 −0.038 0.069 0.053
12 −0.074 −0.065 0.013 0.018 −0.021 −0.087

Downloaded by Marita Nadyah ([email protected])

 lOMoARcPSD|3191521

The residuals are shown below. For both 1974Q4 and 1980Q2 we observe a large
negative residual, followed by a number of large positive residuals in subsequent
quarters. This pattern is typical for the occurrence of an additive outlier in an
AR-model with positive autoregressive coefficients (which is the case here).
3

-1
2
-2
1
-3
0

-1

-2 Residual
Actual
-3 Fitted

1970 1975 1980 1985 1990 1995

Figure 3: Residuals in AR(4) model for quarterly consumption growth rates.

3.1.c) [1 point] Setting the observations for 1974Q4 and 1980Q2 equal to 0, we find
that skewness becomes equal to −0.27, while kurtosis is equal to 3.04. These values
are such that the Jarque-Bera test statistic takes the value 1.44, with a p-value
equal to 0.49 such that normality cannot be rejected. This is also confirmed by
the histogram shown below, which is much more similar to the normal density than
before. Apparently, the two observations in 1974Q4 and 1980Q2 are responsible for
the strong signs of non-normality in the original time series.
.9

.8 Histogram
Normal
.7

.6
Density

.5

.4

.3

.2

.1

.0
-1 0 1 2 3

Figure 4: Histogram of quarterly consumption growth rates, 1970Q1-1999Q4,

with 1974Q4 and 1980Q2 set equal to 0.
The autocorrelations and partial autocorrelations are shown above. We find quite
some differences compared to the estimates obtained before. In addition, the auto-
correlation at lag 8 is no longer significant, while the partial autocorrelations now
are significant for lags 1 to 4, providing a clearer pattern than the partial autocor-
relations for the original series.

Downloaded by Marita Nadyah ([email protected])

 lOMoARcPSD|3191521

3.1.d) [2.5 points] Based on the AR(4) model estimated in 3.1.b., we compute the t-
statistics λ̂i (τ ) for testing whether an additive outlier (AO) or innovation outlier (IO)
occurred at time τ , for τ = 1, . . . , T . These sequences of test statistics are shown
in Figure 5. The largest absolute value of the test statistic for additive outliers is
obtained for 1980Q2, with λ̂AO (τ ) = −5.45. The largest absolute value of the test
statistic for innovation outliers is also obtained for 1980Q2, with λ̂IO (τ ) = −4.83.
For both statistics, the second largest (absolute) value is found for 1974Q4, with
λ̂AO (τ ) = −3.54 and λ̂IO (τ ) = −2.96s.
3 3

2 2

1 1

0 0

-1 -1

-2 -2

-3 -3

-4 -4

-5 -5

-6 -6
1970 1975 1980 1985 1990 1995 1970 1975 1980 1985 1990 1995

Figure 5: Values of the λ̂i (τ ) statistics for testing whether an AO or IO

occurred at time τ are shown on the left and right, respectively.

We use simulation to obtain the (finite-sample) distribution of the test statistics

λ̂AO = max1≤τ ≤T |λ̂AO (τ )| and λ̂IO = max1≤τ ≤T |λ̂IO (τ )| under the null hypothesis
of no outliers, using 10,000 replications. The results of this exercise are shown in
Figure 6. For both test statistics we find very small p-values, equal to 0.000. We
therefore reject the null hypothesis of no outliers. For 1980Q2 (where the maximum
is achieved, as noted above) the value of the λ̂AO (τ ) statistic is somewhat larger (in
absolute value) than the value of the λ̂IO (τ ) statistic. However, the difference is not
very large (while their distributions are very similar) so that it is not completely
obvious whether this observation is best characterized as an AO or as an IO.
1.2 1.2

lambda-hat = 5.45 lambda-hat = 4.83

1.0 p-value = 0.000 1.0 p-value = 0.000

0.8 Histogram 0.8 Histogram

Frequency

Frequency

Extreme-Max Extreme-Max
0.6 0.6
q(0.90) = 3.42 q(0.90) = 3.44
q(0.95) = 3.64 q(0.95) = 3.66
0.4 q(0.99) = 4.04 0.4 q(0.99) = 4.08

0.2 0.2

0.0 0.0
1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0

Figure 6: Simulated distribution under the null hypothesis of λ̂i (τ ) statistics for testing
for the presence of an AO or IO are shown on the left and right, respectively.

Downloaded by Marita Nadyah ([email protected])

 lOMoARcPSD|3191521

3.1.e) [2.5 points] First, we estimate an AR(4) model in which the observations in
1974Q4 and 1980Q2 are treated as AO’s, that is

φ4 (L)(yt − µ − δ1 d1974Q4,t − δ2 d1980Q2,t ) = εt

where φ4 (L) = 1−φ1 L−· · ·−φ4 L4 , and where d1974Q4,t and d1980Q2,t are dummy vari-
ables for 1974Q4 and 1980Q2, respectively. Using the observations for t =1970Q1-
1999Q4, we find the following least squares estimates, with standard errors in paren-
theses: µ̂ = 0.903(0.088), φ̂1 = 0.201(0.092), φ̂2 = 0.292(0.089), φ̂3 = 0.223(0.092),
φ̂4 = −0.300(0.090), δ̂1 = −2.336(0.508), and δ̂2 = −3.400(0.510). Note that all
four AR-coefficients are significantly different from zero at the 5% level in this case.
We find negative values for δ̂1 and δ̂2 as expected, given that consumption growth
experienced large drops in both quarters. The estimates of δ1 and δ2 are significant
even at the 1% level, while they also are large in magnitude (compared to µ̂).
3

-1
2
-2
1
-3
0

-1

-2 Residual
Actual
-3 Fitted

1970 1975 1980 1985 1990 1995

Figure 7: Residuals in AR(4)-AO model for quarterly consumption growth rates.

The R2 of the model is equal to 0.426, which is more than three times as large as
the R2 for the AR(4) model estimated in 3.1.b (the residual variance is reduced
by about 33% from 0.45 to 0.30). The residuals from the AR(4) model with AO’s
are shown above. Comparing this with the ‘standard’ AR(4) model, the main dif-
ferences (obviously) occur in the residuals for 1974Q4 and 1980Q2 and subsequent
observations. In the AR(4)-AO model, the residuals have skewness equal to −0.28
and kurtosis equal to 3.09. Normality is not rejected, as the Jarque-Bera statistics
is equal to 1.56, with a p-value of 0.46. The (partial) autocorrelations of the resid-
uals are shown above; for both, no significant values occur. This also applies to the
(partial) autocorrelations of the squared residuals.

Second, we estimate an AR(4) model in which these observations are treated as IO’s,
that is
φ4 (L)(yt − µ) = δ1∗ D1974Q4,t + δ2∗ D1980Q2,t + εt .
Using the observations for t =1970Q1-1999Q4, we find the following least squares
estimates, with standard errors in parentheses: µ̂ = 0.920(0.083), φ̂1 = 0.157(0.082),
φ̂2 = 0.129(0.081), φ̂3 = 0.186(0.083), φ̂4 = −0.137(0.082), δ̂1 = −2.077(0.616), and
δ̂2 = −3.089(0.607). Note that the estimated coefficients for the second and fourth

Downloaded by Marita Nadyah ([email protected])

 lOMoARcPSD|3191521

lag are not significantly different from zero at the 10% level. The estimates of δ1
and δ2 are negative and significant even at the 1% level, while they also are large in
magnitude (compared to µ̂.

The R2 of the model is equal to 0.347, which is two-and-a-half times as large as

the R2 for the AR(4) model estimated in 3.1.b (the residual variance is reduced by
about 25% from 0.45 to 0.34). The residuals from the AR(4) model with IO’s are
shown below. Comparing this with the ‘standard’ AR(4) model, the main differences
(obviously) occur in the residuals for 1974Q4 and 1980Q2, which are exactly equal
to 0 in the IO-model. In the IO-model, the residuals have skewness equal to −0.24
and kurtosis equal to 3.04. Normality is not rejected, as the Jarque-Bera statistics is
equal to 1.17, with a p-value of 0.56. The (partial) autocorrelations of the residuals
are shown in the table above; for both, no significant values occur. This also applies
to the (partial) autocorrelations of the squared residuals.
3

-1
2
-2
1
-3
0

-1

-2 Residual
Actual
-3 Fitted

1970 1975 1980 1985 1990 1995

Figure 8: Residuals in AR(4)-IO model for quarterly consumption growth rates.

Based on the above findings, it seems that the AR(4)-AO model should be preferred.
This is based on the following arguments:
• In the ‘standard’ AR(4) model, the patterns in the residuals around the obser-
vations in 1974Q4 and 1980Q2 resemble additive outliers more than innovation
outliers.
• The AR(4)-AO model provides the best fit (in terms of R2 and significance of
the AR-coefficients)

Finally, comparing Figure 7, which shows the fitted values and residuals from the
model with outliers, with Figure 3 shows that the largest residuals in 1974Q4 and
1980Q2 obviously have disappeared, but some others remain; for example in 1973Q2.
In fact, all residuals for the four quarters from 1973Q2-1974Q1 are relatively large
and negative. Also around other recession periods (1981, 1990-1), we observe rel-
atively large negative residuals. It may be worthwhile to examine whether some
of these observations also should be considered as outliers, or whether a nonlinear
model (e.g. a threshold model) may be necessary to capture these.

Downloaded by Marita Nadyah ([email protected])

 lOMoARcPSD|3191521

3.1.f) [1 point] We use the three estimated AR(4) models to obtain one-step ahead
forecasts for the quarterly growth rates of consumption for the sample period 2000Q1-
2013Q4.

The three series of forecasts are shown below, together with the actual values of
consumption growth. We observe that the three forecasts are almost identical. The
MSPE’s are equal to 0.180 (AR(4)), 0.196 (AR(4)-AO) and 0.225 (AR(4)-IO). Hence,
incorporating outliers in the model actually leads to less accurate forecasts. All three
models produce biased forecasts, in the sense that the mean forecast errors of −0.171,
−0.207 and −0.249 (for the AR(4), AR(4)-AO and AR(4)-IO models, respectively)
are significantly different from 0 at the 5% significance level. From Figure 9, it
appears that the mean reason for the negative bias are due to 2007-9, for which
all three models give forecasts that are much higher than the actual consumption
growth rate.
2.0

1.5

1.0

0.5

0.0

-0.5
AR(4) forecasts
AR(4)-AO forecasts
-1.0 AR(4)-IO forecasts
Actual growth rate

-1.5
2000 2002 2004 2006 2008 2010 2012

Figure 9: Forecasts for quarterly consumption growth rates, 2000Q1-2013Q4.

In sum, treating the observations in 1974Q4 and 1980Q2 as outliers is not useful
from a forecasting perspective. One final remark to qualify this statement is that
the usefulness of the forecasts seems limited anyway: the variance of the actual
quarterly growth rates over the forecast period is equal to 0.243, such that even the
MSPE for the AR(4) forecasts is only about 25% smaller.

Downloaded by Marita Nadyah ([email protected])