Sei sulla pagina 1di 30

MFEFM TERM PAPER

GROUP - III
Abhay Mittal |
15P061
Aditya Bansal |
15P064
Apurv Gupta |
15P069
Ayan Panda |

Contents
Introduction: What are we trying to achieve?.....................................................2
PART -1: Regression............................................................................................... 3
OBJECTIVE:....................................................................................................... 3
RATIONALE FOR SELECTING INDEPENDENT VARIABLES:..........................3
DATA DESCRIPTION:........................................................................................ 3
MODELLING PROCEDURE:..............................................................................5
OUTPUT:............................................................................................................ 7
PART II: ARIMA and GARCH Modelling...................................................................9
OBJECTIVE ...................................................................................................... 9
DATA DESCRIPTION ...................................................................................... 9
PROCEDURE ................................................................................................... 9
ARCH EFFECT TEST FOR HETEROSKEDASTICITY:......................................18
MODELLING OF ERROR VARIANCE USING EGARCH:.................................19
PART III: VAR AND CO-INTEGRATION...................................................................21
OBJECTIVE:..................................................................................................... 21
PROCEDURE:.................................................................................................. 21
CONCLUSION....................................................................................................... 26
SCOPE FOR IMPROVEMENT..................................................................................27

1 | Page

Introduction: What are we trying to


achieve?
Our model is primarily devised to understand an empirical dependency between
the share price of Exxon Mobil and Shell in addition to Dow Jones index, Brent Oil
price, Gold price and USD/EUR exchange rate. By default, we keep the share
price of Exxon Mobil as a dependent variable while treating all others as
independent variables. The categorical aim of this study is to understand
whether there exists any arbitrage opportunity between these variables or are
they all independent of each other. In either case, we then go ahead to develop a
forecasting model to estimate the share price of Exxon Mobil over a period after
taking all of these variables into consideration. There is an equal chance that the
share price of Exxon Mobil depends on its past values (an attribute which will be
understood

after

auto

correlation

tests).

Once

these

observations

are

established, we move on to understand if these variables have any cointegrating relationships. If there doesnt exist any co-integrating relationships,
this signifies there does exist ample opportunities for an arbitrage and
depending upon the flow of causality, investment choices need to be made.
However, if there do exist co-integrating relationships, it would then mean that
the prices of all the variables under consideration will converge in the long run
and hence we need to take our decision accordingly. So for instance, if Shell
prices are higher as of now rather than Exxon Mobil, it means we need to go long
on Exxon Mobil and short on Shell prices, for in case of co-integrating
relationships, the prices will converge in the long run. This eventually helps us to
devise a prudent portfolio strategy in order to maximize our returns.
We choose these variables while forecasting the share price value of Exxon Mobil
because Exxon Mobil is into oil refining and Shell is its close competitor. Any
changes in the stock price of the latter is supposed to have effect on Exxons
share price if these changes are due to some industry wide phenomena. Further,
Exxon Mobil is listed on NASDAQ and hence Dow Jones Index can have an effect
on it. Moreover, Brent crude oil is the input to these refining operations and
hence can have an effect. Besides, the USD/EUR exchange rate is important for
all global oil trades are done in the USD.
Thus, after making a judicious statistical study of the above variables we will be
able to determine and forecast the future share price of Exxon Mobil. The model
2 | Page

utilizes ARIMA to make the time data series stationary and then goes on to use
the GARCH model in order to understand the long term causality between the
variables.

PART -1: Regression


OBJECTIVE:
The objective of the study is to model the stock price of Exxon Mobil, a US
based global oil and gas major.

RATIONALE FOR SELECTING INDEPENDENT VARIABLES:


Independent Variable
Crude Oil Price (Brent)
Gas Price - Nymex
Gold
Shell Equity
Dow Jones and Nasdaq

Rationale
As reserves are valued on crude price,
basic asset
Similar reasons as above; basic asset
Basic Commodity
Similar company in the same sector
Major stock indices

DATA DESCRIPTION:

Source of Data Bloomberg Terminal Database


Data Span 12th Feb, 2016 15th Sept, 2016
Data Plots

3 | Page

Descriptive Statistics

nasdaq

Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance

nymex-gas
4923.426
028
10.15133
864
4952.251
#N/A
202.7728
306
41116.82
082
0.352549
867
0.742214
392
1017.089
4266.837
5283.926
1964446.
985
399

Kurtosis
Skewness
Range
Minimum
Maximum
Sum
Count

Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance
Kurtosis
Skewness
Range
Minimum
Maximum
Sum
Count

exxoneq

Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance
Kurtosis
Skewness
Range
4 | Page

83.62992
481
0.276114
178
84.22
85.21
5.515376
381
30.41937
663
0.452208
081
0.291729
148
26.41

brent

2.469646
617
0.017843
033
2.591
2.716
0.356414
301
0.127031
154
0.982683
977
0.559092
986
1.377
1.639
3.016

Kurtosis

48.96192
982
0.465390
293
48.58
48.61
9.296163
824
86.41866
184
0.532407
744

Skewness
Range
Minimum
Maximum

0.077115
461
39.89
27.88
67.77

Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance

985.389 Sum
399 Count

shelleq

dowjones

Kurtosis

23.81966
165
0.142222
222
23.32
22.54
2.840886
659
8.070637
008
0.729786
88

Skewness
Range

0.229271
898 Skewness
12.77 Range

Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance

19535.81
399

Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance
Kurtosis

17580.17
922
34.57001
966
17749.31
18533.05
690.5356
02
476839.4
176
0.076752
374
0.925163
434
2975.87

Minimum
Maximum

68.71 Minimum
95.12 Maximum

Sum

33368.34 Sum

Count

399 Count

16.67 Minimum
29.44 Maximum
9504.045 Sum
399 Count

15660.18
18636.05
7014491.
51
399

gold
Mean
Standard Error
Median
Mode
Standard
Deviation
Sample
Variance
Kurtosis
Skewness
Range
Minimum
Maximum
Sum
Count

1196.2237
34
4.1051669
8
1193.57
1272.61
82.000646
2
6724.1059
77
0.8138992
5
0.2590161
9
313.39
1052.94
1366.33
477293.27
399

MODELLING PROCEDURE:

We propose a linear model, we first do linear fit, check DW which tends to


0, so we do LM test, which confirms autocorrelation, then we use lagged
dependent variable (T-1) in the model. remove insignificant variables and

finally narrow down the variable list and obtain coefficients


Iteration 1

5 | Page

DW is very close to

Breus ch-Godfrey Serial Correlation LM Tes t:

0 so we have to

F-statistic
Obs *R-s quared

check

for

correlation

0.0000
0.0000

Tes t Equation:
Dependent Variable: RESID
Method: Leas t Squares
Date: 09/16/16 Time: 10:59
Sample: 2/12/2015 9/15/2016
Included obs ervations : 399
Pres ample mis sing value lagged res iduals s et to zero.
Variable

Coefficient

Std. Error

t-Statistic

C
BRENT
DOWJONES
GOLD
NASDAQ
NYMEX_GAS
SHELLEQ
RESID(-1)

0.598740
0.011801
-7.87E-05
0.000215
0.000165
-0.159234
-0.019678
0.881709

1.200653
0.010340
0.000216
0.000920
0.000594
0.193147
0.037140
0.023927

0.498679
1.141352
-0.364838
0.233659
0.277559
-0.824416
-0.529825
36.84985

0.776432
0.772430
0.821672
263.9818
-483.7468
193.9873
0.000000

Mean dependent var


S.D. dependent var
Akaike info criterion
Schwarz criterion
Hannan-Quinn criter.
Durbin-Watson stat

Prob.
0.6183
0.2544
0.7154
0.8154
0.7815
0.4102
0.5965
0.0000
2.61E-14
1.722428
2.464896
2.544875
2.496572
1.947584

LM Test Result
-

Hence we will now add (t-1) value of exxonmobils equity as a


parameter in regression

Prob. F(1,391)
Prob. Chi-Square(1)

auto-

R-squared
Adjusted R-s quared
S.E. of regres sion
Sum s quared res id
Log likelihood
F-statistic
Prob(F-s tatis tic)

1357.911
309.7965

Iteration 2

6 | Page

Dw is acceptable at
1.73 close to 2
Model is not false as F
prob is 0
R^2 is high, so good
model
T
stat
of
some
parameters
shows
they
are
not
significant
They are shell equity
price and natural gas
price
We first remove shell
equity price, as we
expect natural gas
price to be relevant
as revenue earner

Iteration 3

7 | Page

Dependent Variable: EXXONEQ


Method: Least Squares
Date: 09/16/16 Time: 11:01
Sample (adjusted): 2/13/2015 9/15/2016
Included observations: 398 after adjustments
Variable

Coefficient

Std. Error

t-Statistic

Prob.

C
BRENT
DOWJONES
GOLD
NASDAQ
NYMEX_GAS
SHELLEQ
EXXONEQ(-1)

-1.861678
0.025689
0.002917
0.002409
-0.006330
0.064791
-0.086054
0.754568

1.495739
0.012559
0.000382
0.001175
0.000921
0.235871
0.045138
0.026821

-1.244654
2.045469
7.631310
2.050453
-6.869971
0.274687
-1.906452
28.13340

0.2140
0.0415
0.0000
0.0410
0.0000
0.7837
0.0573
0.0000

R-squared
Adjus ted R-squared
S.E. of regression
Sum squared resid
Log likelihood
F-statis tic
Prob(F-statistic)

0.967806
0.967228
0.996546
387.3104
-559.3197
1674.839
0.000000

Mean dependent var


S.D. dependent var
Akaike info criterion
Schwarz criterion
Hannan-Quinn criter.
Durbin-Watson stat

83.60796
5.504825
2.850853
2.930982
2.882591
1.715999

DW is OK, but brent Dependent Variable: EXXONEQ


Method: Least Squares
now
becomes
nonDate: 09/16/16 Time: 11:07
significant
Sample (adjusted): 2/13/2015 9/15/2016
So we will try to Included observations: 398 after adjustments
remove gas price, and
add

shelleq

to

the

model again as it was


close

to

being

significant

Variable

Coefficient

Std. Error

t-Statistic

Prob.

C
BRENT
DOWJONES
GOLD
NASDAQ
NYMEX_GAS
EXXONEQ(-1)

-2.403467
0.009134
0.002650
0.003426
-0.005708
-0.068997
0.755280

1.473432
0.009104
0.000357
0.001050
0.000865
0.225947
0.026909

-1.631203
1.003361
7.426754
3.261391
-6.601790
-0.305368
28.06827

0.1037
0.3163
0.0000
0.0012
0.0000
0.7602
0.0000

R-squared
Adjusted R-squared
S.E. of regression
Sum squared resid
Log likelihood
F-statistic
Prob(F-statistic)

0.967505
0.967007
0.999898
390.9199
-561.1656
1940.299
0.000000

Mean dependent var


S.D. dependent var
Akaike info criterion
Schwarz criterion
Hannan-Quinn criter.
Durbin-Watson stat

OUTPUT:
Model: Multiple regression frameworks
Y= -2.403 + 0.009*x1 + 0.002*x2 + 0.003*x3 0.006*x4
0.07*x5 + 0.755*Y (-1) + u where u ~ N (0, )
2

Here,
Dependent Variable, Y= Exxon Mobil stock price
Explanatory Variables:
X1 = Brent Crude Prices (US$)
X2 = Dow Jones Index
X3 = Gold Prices (US$)
X4 = Nasdaq Index
X5 = Nymex Gas Price (US$)

8 | Page

83.60796
5.504825
2.855104
2.925217
2.882875
1.722113

*Shell Eq was removed from the final model in the final iteration as T-stat
of the same was not statistically significant.
The above model with an adjusted R2 value of 0.967 implies a sufficient
explanation of the dependent variable in terms of the identified independent
variables.

9 | Page

PART II: ARIMA and GARCH Modelling


OBJECTIVE
Forecasting Brent crude price for the next 7 days based on the historic Brent
crude price series.

DATA DESCRIPTION
We have Brent crude price series from 12/02/15 to 15/09/16 and we are trying to
forecast the price of Brent crude over the next 7 days i.e. From 16/09/16 to
22/09/16.

PROCEDURE
1. Graph of the Original Data series

10 | P a g e

2. Correlogram of the Original Series

Observation - The series is decaying very slowly and there is some hint
about series having a trend which we will confirm in subsequent analysis
3. Checking if the series is stationary or non-stationary
We have used Augmented Dickey Fuller Unit Root Test in order to check for
stationarity both at level and first difference.
3.1 Unit Root Test at Level

11 | P a g e

Observation The probability value from ADF test is 0.749 so we cant reject
the null hypothesis and the series has a unit root at level which means a nonstationary series.
Next we will check for unit root test at first difference series
3.2 Unit Root Test at First Difference

Observation The probability value from ADF test at first difference is 0 so we


can reject the null hypothesis and the series doesnt have a unit root at level
which means it becomes a stationary series on first differencing
4. De trend the series and observe the correlogram of the detrended series

12 | P a g e

Observation The PACF suggests presence of ma(1) in the de trended series


and we will check the model using ar(1) in the least square method

5. Use ma(1) and check for the significance of the model

13 | P a g e

Observation The probability value corresponding to ma(1) is 0.0037 ie. Lesser


than 5% and hence we can say ma(1) is significant in explaining the model.
6. Check for Residual Diagnostics using Q-statistic

Observation We can see none of the levels in Q-statistics Residual Diagnostic


test has a probability value less than 0.05 and therefore the series has reached
White Noise process. We can proceed to forecasting
7. In-sample Forecasting

14 | P a g e

First we will go for in-sample forecasting of the brent crude price using both
static and dynamic forecast to check the mean absolute percentage error in
the model. We will forecast the brent price from 8/09/2016 to 15/09/2016
using the model and compare it with the original data.
7.1 Static In-sample Forecast

Observation The Mean Absolute Percentage error from the static in-sample
forecast is only 2.47% and therefore we can use the model to forecast the future
brent crude prices.

15 | P a g e

But still we will check for dynamic in-sample forecast before we go for future
brent price forecasts.
7.2 Dynamic In-sample Forecast

Observation The Mean Absolute Percentage error from the dynamic in-sample
forecast is only 2.28% and therefore we can use the model to forecast the future
brent crude prices.

16 | P a g e

8. Out of sample Forecasting


Now we will go for out of sample forecasting of the brent crude price using both
static and dynamic forecast. We will forecast the brent price from 16/09/2016 to
22/09/2016.

8.1 Static out of sample Forecast

17 | P a g e

8.2 Dynamic out of sample Forecast

Observation We can see the forecasted data for future brent prices in blue in
the above chart

18 | P a g e

9. Brent Crude Price Forecasts


The forecasted price for Brent crude price over the next 7 days as per the model
is as below:

ARCH EFFECT TEST FOR HETEROSKEDASTICITY:


Univariate ARIMA modelling to forecast the Brent Crude price using OLS
estimation will predict the conditional mean of these series based on past
information with the assumption that the residual variance is constant over
time.
But these series exhibit high variability suggesting that conclusion/forecast
based on OLS is likely to be inaccurate, hence a correction is required to reduce
this variability or volatility or risk or heteroscedastic error variance.

19 | P a g e

20 | P a g e

Observations:
According to the above ARCH effect test the null hypothesis,
H0: 1 = 2 = = q = 0 against
H1: 1 2 q 0,
we can observe that null hypothesis can be rejected and there exists autocorrelation in error variance. As can also be observed, ARCH lag of order (1) is
present herein.

MODELLING OF ERROR VARIANCE USING EGARCH:


EGARCH model is employed for variance of the error term, the same is described
as:

log( ht ) 0 j
j 1

ut j
h t j

j
j 1

ut j
h t j

i log( ht i )
i 1

Here,
0 : the mean of the volatility equation
: represents the size effect, which indicates how much volatility increases
irrespective of the direction of the shock
: represents the sign effect, which examine whether shocks have asymmetric
or symmetric effects on volatility. When < 0, a positive shock u t-j >= 0 (goodnews) generates less volatility than a negative shock (bad-news) u t-j < 0.
: represents an evaluation of the persistence of shocks. Nelson shows that
absolute value of <1 ensures stationarity.
Following output describes the various coefficients in the equation as modelled,

21 | P a g e

As per the above output table, C (5) or is observed to be statistically


significant. However, the coefficient is positive so Leverage Effect is absent,
in other words a positive shock and a negative shock will have symmetric effects
on volatility.
Other coefficients being statistically insignificant can be safely ignored.

PART III: VAR AND CO-INTEGRATION


OBJECTIVE:
To determine whether the Exxon series is co-integrated with others
(shell/crude/gas). Depending upon the test results, what is the nature of
relationship, causality, regression.

PROCEDURE:
We begin with a logarithmic transformation of the series, and then move onto
unit root tests for determining the order of integration as we are certain from
previous sections that these series are non-stationary.
1. ADF Unit Root Tests

22 | P a g e

Null Hypothesis: LOGBRENT has a unit root


Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

t-Statistic

Prob.*

-1.629389

0.7798

t-Statistic

Prob.*

-22.70015

0.0000

t-Statistic

Prob.*

-1.512299

0.8241

t-Statistic

Prob.*

-20.28603

0.0000

t-Statistic

Prob.*

-2.200647

0.4874

t-Statistic

Prob.*

-18.50292

0.0000

Null Hypothesis: DLOG(BRENT) has a unit root


Exogenous: Constant
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

Null Hypothesis: LOGGAS has a unit root


Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

Null Hypothesis: D(LOGGAS) has a unit root


Exogenous: Constant
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

Null Hypothesis: LOGSHELL has a unit root


Exogenous: Constant, Linear Trend
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

Null Hypothesis: D(LOGSHELL) has a unit root


Exogenous: Constant
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

Null Hypothesis: LOGEXON has a unit root


Exogenous: Constant, Linear Trend

23 | P a g e

Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

t-Statistic

Prob.*

-2.627722

0.2682

t-Statistic

Prob.*

-20.26300

0.0000

Null Hypothesis: D(LOGEXON) has a unit root


Exogenous: Constant
Lag Length: 0 (Automatic - based on SIC, maxlag=16)

Augmented Dickey-Fuller test statistic

Observation:
The above results show that the log series of Exxon, shell, Brent & Gas are I (1),
and hence we should proceed to check for cointegration among them.
2. Lag Structure
The lag structure as computed using the Lag Length criteria:

Observation:
As per above output, lag of order 2 is found to be optimum.
3. Johanssen Juselius Cointegration Test
Below table output uses the Maximum Eigen Value Test as well as the Trace
statistics to conclude if there exists co-integration amongst the series.
Sample (adjusted): 2/18/2015 9/15/2016
Included observations: 396 after adjustments
Trend assumption: Linear deterministic trend
Series: LOGEXXON LOGSHELL LOGBRENT LOGGAS
Lags interval (in first differences): 1 to 2

24 | P a g e

Unrestricted Cointegration Rank Test (Trace)


Hypothesized
No. of CE(s)

Eigenvalue

Trace
Statistic

0.05
Critical Value

Prob.**

None
At most 1
At most 2
At most 3

0.041020
0.024608
0.014024
0.004805

33.95333
17.36684
7.500103
1.907411

47.85613
29.79707
15.49471
3.841466

0.5044
0.6129
0.5203
0.1672

Trace test indicates no cointegration at the 0.05 level


* denotes rejection of the hypothesis at the 0.05 level
**MacKinnon-Haug-Michelis (1999) p-values
Unrestricted Cointegration Rank Test (Maximum Eigenvalue)
Hypothesized
No. of CE(s)

Eigenvalue

Max-Eigen
Statistic

0.05
Critical Value

Prob.**

None
At most 1
At most 2
At most 3

0.041020
0.024608
0.014024
0.004805

16.58650
9.866733
5.592692
1.907411

27.58434
21.13162
14.26460
3.841466

0.6156
0.7571
0.6658
0.1672

Max-eigenvalue test indicates no cointegration at the 0.05 level

Observation:

As none of the ranks in both the tests are statistically significant, it can be
concluded that there exists no co-integration amongst these price series.
Hence, now we will proceed to Unrestricted VAR modelling and Granger
causality determination in the short run.

4. Vector Auto Regression Results


Sample (adjusted): 2/17/2015 9/15/2016
Included observations: 397 after adjustments
Standard errors in ( ) & t-statistics in [ ]
LOGEXXON

LOGSHELL

LOGBRENT

LOGGAS

LOGEXXON(-1)

0.988093
(0.06222)
[ 15.8817]

0.315485
(0.08744)
[ 3.60804]

0.047889
(0.12922)
[ 0.37059]

-0.260079
(0.12767)
[-2.03717]

LOGEXXON(-2)

-0.006742
(0.06180)
[-0.10909]

-0.308540
(0.08686)
[-3.55215]

-0.044774
(0.12837)
[-0.34880]

0.287893
(0.12682)
[ 2.27009]

LOGSHELL(-1)

0.020837
(0.04478)

0.958947
(0.06293)

0.123467
(0.09300)

0.093782
(0.09188)

25 | P a g e

[ 0.46536]

[ 15.2382]

[ 1.32757]

[ 1.02068]

LOGSHELL(-2)

-0.047348
(0.04481)
[-1.05674]

-0.000794
(0.06297)
[-0.01261]

-0.124469
(0.09306)
[-1.33748]

-0.131493
(0.09194)
[-1.43020]

LOGBRENT(-1)

-0.045720
(0.03093)
[-1.47810]

-0.021472
(0.04347)
[-0.49392]

0.803608
(0.06425)
[ 12.5085]

-0.041435
(0.06347)
[-0.65282]

LOGBRENT(-2)

0.061607
(0.03096)
[ 1.98977]

0.040172
(0.04351)
[ 0.92318]

0.194979
(0.06431)
[ 3.03193]

0.085966
(0.06353)
[ 1.35308]

LOGGAS(-1)

-0.003531
(0.02489)
[-0.14189]

0.028101
(0.03498)
[ 0.80342]

0.030051
(0.05169)
[ 0.58134]

0.968453
(0.05107)
[ 18.9637]

LOGGAS(-2)

-0.005165
(0.02482)
[-0.20812]

-0.030552
(0.03488)
[-0.87590]

-0.048102
(0.05155)
[-0.93314]

-0.010887
(0.05093)
[-0.21378]

0.112372
(0.04654)
[ 2.41457]

0.030693
(0.06541)
[ 0.46926]

0.010255
(0.09666)
[ 0.10609]

-0.138236
(0.09550)
[-1.44753]

0.959299
0.958460
0.071533
0.013578
1143.125
1148.055
-5.738311
-5.647995
4.423653
0.066620

0.974672
0.974150
0.141294
0.019083
1866.374
1012.940
-5.057632
-4.967317
3.162443
0.118690

0.979681
0.979262
0.308596
0.028202
2338.470
857.8737
-4.276442
-4.186126
3.871327
0.195839

0.967483
0.966813
0.301205
0.027862
1443.039
862.6860
-4.300685
-4.210369
0.892260
0.152943

R-squared
Adj. R-squared
Sum sq. resids
S.E. equation
F-statistic
Log likelihood
Akaike AIC
Schwarz SC
Mean dependent
S.D. dependent

Determinant resid covariance (dof adj.)


Determinant resid covariance
Log likelihood
Akaike information criterion
Schwarz criterion

1.80E-14
1.64E-14
4047.315
-20.20814
-19.84687

Equations
So, as per VAR, the equations are (truncated to 2 decimals):
logexxon = 0.99*logexxon(-1) + 0.06*logbrent(-2) + 0.11
logshell = 0.32*logexxon(-1) - 0.30*logexxon(-2) + 0.95*logshell(-1)
logbrent = 0.047*logexxon(-1) + 0.80*logbrent(-1) + 0.19*logbrent(-2)
loggas = -0.26*logexxon(-1) + 0.29*logexxon(-2) + 0.97*loggas(-1)
5. Pairwise Granger Causality Tests

26 | P a g e

We conduct a pairwise Granger causality test on the differenced logarithms of


Exxon stock, Shell stock, Brent crude and Nymex gas. The output from Eviews is
as follows:

At the 5% significance level, only dlog(exxon) Granger causes dlog(Shell),


which is along expected lines because Exxon Mobil and Shell are both

giants, but it is surprising that the directionality is not bi-directional


At the 10% level, Exxon Granger causes gas prices, unidirectionally, which
is plausible, but not very probable

CONCLUSION
The above modelling exercise has brought to light the following points:
1. The extent of dependence of exxon mobil stock price upon various parameters
like shell stock price, brent, exchange rate etc.
2. The Brent prices forecast using ARIMA and GARCH models
3. The development of a portfolio strategy for an investor in oil sector by
checking whether exxon mobil stock, shell stock and brent crude offer a benefit
of portfolio diversification.
As there exists no co-integration so it is advisable to invest in all the oil stocks to
have diversified portfolio!

27 | P a g e

SCOPE FOR IMPROVEMENT


The possible limitations as identified are:
1.

Only 400 points of daily data have been used, from the years 2015 to

2016, thus multi-year spanning trends will be missed


2.

We have chosen only 2 oil sector stocks and thus might be missing out on

probable relationships between stocks of other companies


If we take into account the above two factors and model subsequently on the
new data, the results could improve and give a more detailed picture and
comprehensive explanation of the markets.

28 | P a g e

29 | P a g e

Potrebbero piacerti anche