Académique Documents
Professionnel Documents
Culture Documents
Table
of
Contents
Introduction
...............................................................................................................................................
3
The
Theory
.................................................................................................................................................
5
West
Texas
Intermediate
(WTI)
........................................................................................................................................................................
5
10-year
US
Bond
Yield
...........................................................................................................................................................................................
6
Dow
Jones
Industrial
Average
............................................................................................................................................................................
7
Exxon
Mobil
Share
Price
.......................................................................................................................................................................................
8
GBP/USD
Exchange
Rate
......................................................................................................................................................................................
9
US
Unemployment
Rate
.....................................................................................................................................................................................
10
Data
Methodology
....................................................................................................................................
11
Empirical
results
........................................................................................................................................
12
Testing
for
Stationarity
Levels
.....................................................................................................................................................................
12
Unit
root
tests
.........................................................................................................................................................................................................
13
Testing
for
Stationarity
First
Difference
.................................................................................................................................................
15
Unit
root
tests
.........................................................................................................................................................................................................
16
ARIMA
Model
...........................................................................................................................................
17
ARIMA
Forecast
.....................................................................................................................................................................................................
19
Cointegration
Model
.................................................................................................................................
21
Engle-Granger
Test
..............................................................................................................................................................................................
21
Johansen
Test
..........................................................................................................................................................................................................
22
VAR
Lag
Selection
..................................................................................................................................................................................................
22
Vector
Error
Correction
Model
.......................................................................................................................................................................
24
VECM
Output
............................................................................................................................................................................................................
25
VECM
Holdout
Period
Forecast
........................................................................................................................................................................
27
Comparison
of
ARIMA
Model
and
VECM
...................................................................................................
28
Out
of
Sample
Forecasts
............................................................................................................................
30
Conclusion
................................................................................................................................................
32
Appendix
..................................................................................................................................................
34
Appendix
1
...............................................................................................................................................................................................................
34
Appendix
2
...............................................................................................................................................................................................................
35
ADF
Brent
Levels
....................................................................................................................................................................................................
35
ADF-GLS
Brent
Levels
...........................................................................................................................................................................................
35
KPSS
Brent
Levels
...................................................................................................................................................................................................
35
Fractional
Integration
Brent
Levels
..............................................................................................................................................................
35
ADF
US
10
Year
Bond
Levels
.............................................................................................................................................................................
36
ADF-GLS
US
10
Year
Levels
................................................................................................................................................................................
36
KPSS
US
10
Year
Levels
........................................................................................................................................................................................
36
Appendix
3
...............................................................................................................................................................................................................
37
ADF
DJIA
First
Differences
..................................................................................................................................................................................
37
ADF-GLS
DJIA
First
Differences
........................................................................................................................................................................
37
KPSS
DJIA
First
Differences
................................................................................................................................................................................
37
Fractional
Integration
DJIA
First
Differences
............................................................................................................................................
37
ADF
WTI
First
Differences
..................................................................................................................................................................................
38
ADF-GLS
WTI
First
Differences
........................................................................................................................................................................
38
KPSS
WTI
First
Differences
................................................................................................................................................................................
38
Appendix
4
...............................................................................................................................................................................................................
39
ARIMA
(0,1,0)
Output
...........................................................................................................................................................................................
39
ARIMA
(1,1,0)
Output
...........................................................................................................................................................................................
39
Bibliography
..............................................................................................................................................
40
Introduction
For
our
research
project,
we
chose
to
try
to
develop
a
model
to
forecast
the
price
of
Brent
crude
oil.
Oil
is
a
commodity
that
is
vital
for
all
economies
for
growth,
prosperity
and
convenience.
For
centuries
before
the
first
oil
well,
people
around
the
world
gathered
oil
around
"seeps,"
places
where
oil
naturally
rose
to
the
surface
and
came
out
of
the
ground
(Yergin
1991).
Collecting
oil
in
that
manner
meant
that
even
productive
areas
did
not
yield
large
amounts
of
oil.
Oil
was
heavily
demanded
in
those
days
for
lighting
as
whale
oil
had
become
too
expensive.
This
meant
that
the
person
who
could
somehow
extract
the
oil
could
capture
the
huge
kerosene
market
(the
refined
product
of
crude
oil
used
in
lamps).
In
1859
the
first
known
oil
well
was
drilled,
using
similar
techniques
as
those
used
in
salt
exploration.
This
was
(eventually)
a
success,
beginning
the
start
of
a
new
age
the
oil
age.
In
the
present
day,
the
world
is
dependent
on
oil
for
power
generation
and
for
use
in
many
products.
This
makes
oil
an
extremely
useful
and
powerful
commodity
in
the
past
we
have
seen
many
wars
over
it
(Yergin
1991).
Our
interest
in
predicting
the
price
of
Brent
crude
oil
stems
from
this
desire
around
the
world
to
find
and
extract
oil
in
order
to
use
it
for
development
and
increasing
standard
of
living.
Crude
oil
is
a
vital
commodity
and
is
used
in
many
goods,
for
example
its
by-products
after
refining
are
lubricating
oils,
diesel,
kerosene,
naphtha
(for
use
in
plastics,
medicines,
de-icers,
perfume,
fertilisers),
gasoline/petrol,
butane
and
propane.
This
makes
crude
oil
a
vital
global
commodity.
This
is
evident
in
the
market,
as
over
two
million
barrels
of
Brent
crude
are
traded
per
day
(Smith
2012).
Also
from
Figure
1
it
can
be
seen
that
global
demand
will
increase,
as
non-OECD
countries
develop.
Figure
2
gives
us
a
clear
visual
representation
of
this
huge
growth
in
demand
over
only
the
past
three
years.
This
makes
a
forecasting
model
to
predict
future
price
movements
very
interesting
to
construct
and,
if
accurate,
extremely
useful.
Figure
2
-
Global
Oil
Demand
2010-2013
Map
(Nelder
2012)
In
order
to
produce
a
forecast
for
the
price
of
Brent,
we
aim
to
use
two
modelling
techniques.
We
will
firstly
try
to
use
ARIMA
forecasting
followed
by
a
more
sophisticated
modelling
technique
such
as
cointegration,
vector
autoregression
or
OLS
regression
modelling1.
We
selected
numerous
variables
in
order
to
help
us
build
our
model
which
are
described
below
in
the
theory
section.
The
overall
aim
of
this
project
is
to
determine
which
of
the
variables
we
test
are
important
in
the
relationship
to
Brent
crude
and
to
better
understand
the
statistical
relationships
between
the
variables
and
Brent
crude
prices.
1
Model
choice
will
be
dependent
upon
characteristics
of
the
dataset.
The
Theory
There
are
multiple
factors
that
affect
the
price
of
Brent
crude
oil.
We
decided
to
look
at
the
theoretical
relationship
between
the
price
of
Brent
and
six
other
variables
in
order
to
compare
theory
with
our
future
model,
to
ensure
that
the
model
agrees
with
theory:
1. West
Texas
Intermediate
(WTI)
Crude
Oil
2. 10-year
US
Bond
Yield
3. Dow
Jones
Industrial
Average
4. Exxon
Mobil
Share
Price
5. GBP/USD
Exchange
Rate
6. US
Unemployment
Rate
Over
short-term
periods
the
US
10-year
bond
yields
(not
prices,
prices
move
inversely
to
yield)
should
move
in
tandem
with
Brent
crude.
The
theory
behind
this
is
that
Treasury
yields
depress
when
there
is
a
lot
of
demand
for
Treasury
products,
which
are
considered
ultra-safe
investments.
When
the
global
economy
is
doing
badly,
people
look
for
a
flight
to
quality
and
US
government
bonds
are
often
demanded,
as
well
as
other
instruments
(for
example
Gold)
to
maintain
a
return.
The
reason
that
the
10-year
bond
was
chosen
as
opposed
to
other
US
instruments
(also
perceived
as
risk-free)
is
because
it
is
the
benchmark
rate
that
guides
almost
all
other
interest
rates
(Durden
2013).
Brent
crude
should
also
reduce
in
price
during
downturns
in
the
economy
as
there
will
be
less
demand
due
to
less
demand
for
consumer
goods.
This
can
be
seen
in
Figure
4;
where
in
2007/08
the
financial
crisis
hit,
and
Brent
crude
prices
collapsed
and
US
10-year
bond
yields
dropped.
Over
the
long-term
Brent
shows
a
positive
trend,
whereas
US
10-year
bond
yields
have
a
negative
trend.
This
doesnt
follow
the
theory
above
for
the
short-term,
however
due
to
excess
demand
with
limited
supply
oil
prices
have
generally
increased,
whilst
at
the
same
time
the
US
has
been
strengthening
and
this
is
reflected
in
depressed
yields
over
time.
The
Dow
Jones
Industrial
Average
(DJIA),
The
Dow,
is
a
price-weighted
average
of
30
significant
stocks
traded
on
the
New
York
Stock
Exchange
and
the
NASDAQ.
This
means
that
when
companies
are
doing
well,
the
DJIA
should
increase,
and
this
should
increase
the
demand
for
oil
as
factories
produce
and
the
cycle
goes
on.
This
can
be
seen
in
Figure
5
as
the
DJIA
increases
from
2003-2007,
then
the
DJIA
crashes
and
Brent
crude
prices
collapse
too.
This
pattern
tends
to
hold
over
the
long-term
as
the
two
increase
and
decrease
together
and
seem
to
be
quite
correlated.
The
reason
the
DJIA
was
chosen
over
its
competitors
is
because
it
has
stood
the
test
of
time
and
is
not
considered
to
be
volatile
or
risky
something
we
felt
was
important
due
to
the
increase
in
volatility
in
the
stock
markets
over
the
financial
crisis.
Exxon
Mobil
is
a
multinational
oil
and
gas
company.
It
is
the
world's
third
largest
company
by
revenue
(DeCarlo
2013),
and
also
the
second
largest
publicly
traded
company
by
market
capitalisation
(Associated
Press
2013).
The
company
was
ranked
No.
5
globally
in
Forbes
Global
2000
list
in
2013
(DeCarlo
2013).
With
reserves
of
72
billion
barrels
of
oil
equivalent
at
the
end
of
2007
and
being
the
largest
refiner
in
the
world
(Exxon
Mobil
Annual
Statements),
we
felt
that
this
would
be
a
suitable
company
to
use
in
our
model.
The
link
between
Exxon
Mobils
share
price
and
the
price
of
Brent
crude
should
move
in
tandem
as
their
business
is
based
on
the
commodity.
This
can
be
seen
in
Figure
6,
as
the
two
tend
to
track
each
other,
with
Brent
crude
being
more
volatile
than
Exxon
Mobils
share
price.
This
exchange
rate
is
commonly
referred
to
as
Cable
due
to
the
fact
that
in
the
1800s,
the
dollar/pound
sterling
exchange
rate
was
transmitted
via
transatlantic
cable.
This
is
the
number
of
dollars
per
pound
sterling.
This
is
used
as
a
variable
to
help
forecast
oil
prices
as
oil,
priced
in
dollars,
tends
to
move
higher
as
the
U.S.
dollar
falls,
and
vice-versa.
Its
a
result
of
oil
traders
trying
to
maintain
their
purchasing
power
in
the
event
of
a
weaker
dollar
(Lazzaro
2013).
This
means
that
when
the
dollar
value
falls
(so
that
GBP/USD
increases),
the
oil
prices
should
increase.
This
is
seen
in
general
over
Figure
7;
however
the
financial
crisis
distorts
this
relationship
due
to
the
flight
to
quality
previously
discussed.
We
wanted
to
use
a
highly
traded
currency
such
as
the
Yen,
Euro,
Sterling
or
the
US
dollar.
Due
to
the
Japanese
QE
programmes,
as
well
as
the
Eurozone
Crisis,
we
decided
to
use
Sterling
as
the
US
dollar
counterpart.
Furthermore,
due
to
our
national
interests
and
as
users
of
the
pound
we
were
interested
in
analysing
the
GBP/USD
exchange
rate.
US Unemployment Rate
The
US
unemployment
rate
is
a
key
indicator
used
to
measure
the
state
of
the
US
economy.
This
variable
was
chosen
as
during
slumps
in
the
economy
the
unemployment
rate
should
increase,
which
then
means
that
fewer
goods
are
demanded,
so
companies
produce
less,
leading
to
a
lower
demand
for
Brent
crude
oil.
This
relationship
can
be
seen
around
2007/08
during
the
financial
crisis
as
unemployment
spikes
up,
and
Brent
prices
collapse.
Also
generally
across
the
time
series,
the
two
seem
to
have
an
inverse
relationship.
Overall,
we
chose
our
variables
individually
because
of
the
reasons
outlined
above.
Additionally,
as
you
may
have
noticed,
all
of
our
variables
were
related
to
the
US,
as
Brent
crude
oil
is
traded
in
US
dollars.
However,
it
is
nave
to
think
that
these
are
the
only
related
variables,
as
our
variables
are
dependent
on
multiple
factors
and
are
not
simply
impacted
by
one
another.
For
example
the
Dow
will
be
impacted
by
each
of
the
variables
above
as
well
as
many
more,
such
as
inflation
and
tapering
not
just
Brent
prices.
10
Data
Methodology
We
retrieved
our
historical
data
for
the
price
of
Brent
Crude
oil
from
the
US
Energy
Information
Administration
(www.eia.gov
n.d.);
a
reliable
resource
for
this
commodity.
For
the
additional
explanatory
variables
we
ensured
that
each
historical
dataset
was
equally
from
a
credible
source,
to
allow
our
model
to
be
as
representative
as
possible
(See
Appendix
1
for
full
list
of
sources).
Due
to
our
intention
to
forecast
using
time
series
modelling,
monthly
data
was
to
be
used.
Our
time
period
covered
January
2000
to
August
2013;
giving
us
164
data
points.
A
holdout
period
was
included
over
the
last
12
points
of
the
data,
one
year,
in
our
sample
to
enable
us
to
compare
the
accuracy
of
the
forecast
models.
We
chose
to
select
this
time
period
because
we
wanted
there
to
be
enough
recovery
from
the
financial
crisis
to
be
present
in
our
historical
data,
to
avoid
any
unnecessarily
obscure
results.
We
felt
that
13
years
of
data
should
provide
a
long
enough
sample
to
capture
trends
and
would
provide
a
balanced
data
set
either
side
of
the
financial
crisis
of
2007/08
(The
Economist
2013).
Throughout
this
report
we
will
build
two
models
and
compare
their
accuracy
within
the
holdout
period
chosen.
Having
assessed
each
models
ability
to
forecast
over
the
holdout
period,
testing
out
of
sample
is
where
our
discussion
will
then
lead.
Following
this
we
will
discuss
the
implications
of
each
model
and
its
suitability
in
the
real
word
by
comparing
the
results
to
theory.
11
Empirical
results
Testing
for
Stationarity
Levels
The
first
step
in
any
forecasting
data
analysis
is
to
plot
the
data
so
visual
analysis
can
be
done
to
estimate
which
modelling
technique
could
be
used
and
to
check
for
any
anomalies.
In
this
case
we
need
to
check
the
individual
data
plots
for
any
visual
signs
of
stationarity.
This
would
be
shown
by
a
constant
mean
and
variance
that
is
independent
of
time.
From
these
plots
(Figure
9)
it
looks
as
though
each
of
the
variables
is
non-stationary
with
each
showing
signs
of
a
trend
(maybe
GBP/USD
does
not
show
a
trend)
and
a
non-constant
variance.
However,
this
is
only
a
visual
analysis.
To
be
sure
we
need
to
undertake
statistical
tests
to
confirm
our
preliminary
thoughts.
12
Statistical
testing
for
stationarity
needs
to
be
done
before
being
able
to
build
any
of
our
forecasting
models
through
unit
root
and
stationarity
tests,
as
we
cannot
simply
rely
on
our
visual
analysis.
ADF
and
ADF-GLS
unit
root
tests
examine
the
null
hypothesis
that
the
data
set
contains
a
unit
root.
The
ADF
test
does
not
hold
much
power
if
there
are
shifts
in
either
the
constant
or
trend
in
the
data;
which
is
why
we
further
test
with
the
ADF-GLS
test.
This
latter
test
deals
with
the
deterministics
in
a
more
sophisticated
manner,
resulting
in
a
more
reliable
inference.
Following
on
from
conducting
the
unit
root
tests
we
can
use
the
complementary
KPSS
test
to
directly
see
if
the
series
are
stationary
around
a
constant
or
trend.
The
null
hypothesis
for
the
KPSS
test
is
that
the
data
series
is
stationary.
In
order
for
the
data
series
to
be
used
within
our
forecasting
models
they
need
to
hold
certain
characteristics
in
their
levels.
The
null
from
the
ADF
and
ADF-GLS
would
need
to
not
be
rejected,
as
we
need
the
series
to
be
I(1)
and
to
contain
a
unit
root.
Whilst
the
KPSS
null
hypothesis
should
be
rejected
to
further
suggest
that
the
series
is
non-stationary.
Additionally,
if
there
are
any
discrepancies
in
the
inference
drawn
from
the
unit
root
and
stationarity
tests,
we
will
need
to
use
fractional
integration
to
decipher
the
true
inference.
Fractional
integration
would
need
to
be
used
on
variables
where
its
series
may
be
stationary;
however
it
takes
a
long
time
to
revert
to
its
mean,
therefore
the
correct
inference
may
not
be
picked
up
in
the
earlier
ADF/ADF-GLS
and
KPSS
tests.
This
contains
two
tests;
the
Whittle
estimator
and
the
GPH
estimator,
which
provide
estimates
for
the
d.
Further,
the
null
hypotheses
are
that
d=0,
therefore
if
we
reject
this,
the
series
are
non-stationary,
I(1).
13
We
used
a
lag
of
6
in
both
the
ADF
and
ADF-GLS2
tests,
but
a
lag
of
4
in
the
KPSS
test3
and
then
let
GRETL
set
it
for
us
in
the
test
for
fractional
integration.
The
5%
level
of
significance
was
used
to
decide
whether
the
null
hypotheses
in
the
ADF
test
should
be
rejected
or
not;
in
our
case
we
did
not
reject
the
null
for
any
level
of
the
variable.
With
the
ADF-GLS
test
at
the
5%
level
of
significance
its
critical
value
was
-2.93,
therefore
we
do
not
reject
the
null
for
all
variables
which
have
t-statistics
larger
than
-2.93.
The
KPSS
test
has
a
critical
value
of
0.148
at
the
5%
level
of
significance;
therefore
all
variables
with
t-
statistics
larger
than
this
value
should
be
rejected,
indicating
a
non-stationary
series.
After
completing
fractional
integration
for
Brent
and
WTI,
we
can
say
all
series
are
non-stationary,
I(1),
and
will
need
to
be
differenced
to
be
made
stationary
(see
Figure
10)
(see
Appendix
2
for
a
sample
of
our
test
results).
Variable
Brent
US/10/year/Bond/Yield
Dow/Jones/Industrial/Average
Exxon/Mobil
GBP/USD/FX/Rate
US/Unemployment/Rate
WTI
ADF?5%?significance?level
P)value
Stationary?
with/constant/
0.6095
NO
0.7328
NO
0.3352
NO
0.8922
NO
0.4177
NO
0.4056
NO
0.7302
NO
ADF(GLS?5%?critical?value?=?(2.93
T)stats/with/constant/
Stationary?
and/trend/
(3.47632
YES
(2.44434
NO
(2.07014
NO
(1.97456
NO
(1.71309
NO
(2.40142
NO
(4.22819
YES
KPSS?5%?critical?value?=?0.148
T)stats
Stationary?
with/trend/
0.0958324
YES
0.277376
NO
0.141104
NO
0.238685
NO
0.563796
NO
0.379412
NO
0.111855
YES
Overall/
stationary?
Conflict
NO
NO
NO
NO
NO
Conflict
Figure
10
-
Unit
Root
and
Stationarity
Tests
of
the
Levels
14
To
be
used
in
an
ARIMA
model,
the
variables
have
to
be
stationary,
I(0),
in
the
first
difference.
So
again,
here
we
need
to
plot
the
data,
but
this
time
in
its
first
difference.
When
looking
at
the
differenced
series
they
all
appear
to
have
stationary
characteristics
-
constant
mean
and
variance
whilst
being
independent
of
time.
Figure
11
-
Data
Plots
of
the
First
Differences
of
the
Series
15
We
followed
the
same
method
as
with
the
levels
of
series
and
have
found
that
all
series
are
stationary
in
their
first
difference
as
shown
in
the
table
below
(Figure
12).
Variable
Brent
US0100year0Bond0Yield
Dow0Jones0Industrial0Average
Exxon0Mobil
GBP/USD0FX0Rate
US0Unemployment0Rate
WTI
ADF=5%=significance=level
P)value
Stationary?
without0constant0
1.80E&08
YES
5.12E&10
YES
2.83E&06
YES
7.28E&137
YES
9.49E&29
YES
0.00713
YES
3.80E&10
YES
ADF&GLS=5%=significance=level
P)value0with0
Stationary?
constant
2.94E&13
YES
8.94E&18
YES
0.18440
NO
0.06439
NO
0.00479
YES
0.01218
YES
7.98E&09
YES
KPSS=5%=critical=value===0.464
T)stats
Stationary?
no0trend
0.03541
YES
0.07187
YES
0.12026
YES
0.08878
YES
0.10722
YES
0.17270
YES
0.02564
YES
Overall0
stationary?
YES
YES
Conflict
Conflict
YES
YES
YES
Figure
12
-
Unit
Root
and
Stationarity
Tests
of
the
First
Differences
These
initial
tests
have
enabled
us
to
decide
on
the
relevant
forecasting
models
that
we
could
create
to
enable
us
to
gain
the
most
accurate
forecast
of
the
price
of
Brent
Crude
oil.
Our
results
imply
that
our
variables
are
integrated
of
the
same
order,
1,
and
can
therefore
be
used
for
a
cointegration
analysis,
as
well
as
an
ARIMA
forecast
as
they
are
stationary
in
their
first
difference
(see
Appendix
3
for
a
sample
of
our
test
results).
16
ARIMA
Model
From
completing
our
unit
root
and
stationarity
tests
we
now
know
that
the
level
of
Brent
has
to
be
differenced
once
to
be
made
stationary4
therefore
we
can
use
it
in
an
ARIMA
forecasting
model.
In
this
ARIMA
model
the
price
of
Brent
Crude
oil
will
be
forecasted
based
upon
its
historical
values.
The
accuracy
of
the
model
will
be
dependent
upon
our
use
of
AR
and
MA
terms
in
its
structure.
As
we
are
trying
to
model
the
price
of
a
volatile
commodity,
we
may
find
that
the
best
model
will
follow
a
random
walk.
Before
testing
what
combinations
of
AR
and
MA
terms
provide
us
with
the
most
accurate
model,
we
first
need
to
analyse
the
correlogram
of
the
first
difference
of
Brent
when
it
is
in
its
stationary
state;
d_Brent.
Figure
13
-
Correlogram
for
the
Differenced
Series
of
Brent,
d_Brent
4
To
forecast
using
an
ARIMA
model
the
variable
needs
to
be
stationary.
17
When
visually
inspecting
the
correlograms,
we
are
looking
to
see
if
there
is
any
pattern
in
either
of
the
ACF
or
PACF,
in
addition
to
observing
if
there
are
any
spikes
at
lags
which
are
significantly
different
from
zero
at
the
5%
level
of
significance5.
From
our
correlogram
we
can
see
that
there
is
potentially
a
slight
pattern
in
the
ACF
and
a
significant
spike
at
lag
1
of
the
PACF
at
the
5%
significance
level,
suggesting
that
there
may
be
a
possible
AR
term
in
the
ARIMA
model.
Following
this
visual
analysis
we
begin
to
test
our
models,
starting
by
setting
our
p
term
to
1,
to
capture
the
AR,
and
the
d
to
1
as
we
need
to
work
with
the
differenced
series.
There
is
no
visual
evidence
of
an
MA
term
as
there
is
no
pattern
in
the
PACF,
so
the
q
term
is
set
to
zero.
When
analysing
the
statistical
output
for
the
ARIMA
(1,1,0)
model
we
can
see
that
the
phi
term
is
significantly
different
from
zero
at
the
5%
level
of
significance
as
5.27e-07
is
less
than
0.05
(see
Appendix
4)
leading
us
to
reject
the
null
hypothesis
that
it
is
equal
to
zero.
This
implies
that
there
is
some
AR
relationship
in
the
data.
Following
finding
this
significant
AR
term
we
decided
to
test
to
see
if
there
was
a
second,
however
this
second
phi
term
appeared
to
be
insignificantly
different
from
zero
at
the
5%
level
and
we
were
able
to
conclude
that
we
only
wanted
to
include
the
1
AR
term
in
our
ARIMA
model.
Using
our
intuition
and
belief
that
the
forecasting
of
Brent
Crude
oil
could
be
represented
by
a
random
walk
ARIMA
model
we
decided
to
create
the
ARIMA
(0,1,0)
model.
In
this
case
we
are
saying
that
there
is
no
AR
or
MA
relationship
and
that
the
best
forecast
of
its
price
in
the
next
period,
is
its
price
in
the
last
period
plus
an
error
term.
In
order
to
compare
these
two
models
capability
of
forecasting
we
looked
at
the
Akaike
criterion
(AIC)
and
their
forecast
evaluation
statistics.
5
These
will
show
as
they
will
be
outside
of
the
confidence
intervals.
18
ARIMA Forecast
Forecast&Evalution&Statistics
ARIMA&(1,1,0)
ARIMA&(0,1,0)
Akaike&criterion&(AIC)
945.6751
964.887
Mean&Error&&&&&&&&&&&&&&&&&&&&&&&
Mean&Squared&Error&&&&&&&&&&&&&&&&
Root&Mean&Squared&Error&&&&&&&&&&
Mean&Absolute&Error&&&&&&&&&&&&&&
Mean&Percentage&Error&&&&&&&&&
Mean&Absolute&Percentage&Error&&&
Theil's&U&&&&&&&&&&&&&&&&&&&&&&&&&
Bias&proportion,&UM&&&&&&&&&&&&
Regression&proportion,&UR&&&&&&
Disturbance&proportion,&UD&&&&&&
:14.371
242.71
15.579
14.371
:13.412
:8.18
97.333
9.8657
8.18
:7.7096
7.7096
2.8027
0.68745
0.16811
0.14444
13.412
4.3741
0.85095
0.091117
0.057933
When
comparing
the
AIC
from
both
models
we
can
see
that
as
the
ARIMA
(1,1,0)
AIC
is
lower
at
964.887,
we
would
expect
it
to
be
a
better
model
as
it
contains
an
AR
term.
However,
when
considering
the
statistical
output
from
the
forecasting
models,
we
can
see
that
the
random
walk
ARIMA
(0,1,0)
creates
more
accurate
forecasts
because
of
the
lower
Theils
U,
and
forecasting
error
statistics
(Figure
16).
As
we
are
testing
two
models
on
their
ability
to
forecast,
we
will
choose
the
random
walk
model
of
ARIMA
(0,1,0)
to
compare
to
our
other
19
model6.
The
forecasts
from
the
ARIMA
(0,1,0)
are
shown
in
Figure
14.
This
shows
that
the
forecast
predicts
a
steady
increase
in
the
price
of
Brent
crude
over
the
holdout
period.
Each
observation
is
within
the
95%
confidence
intervals
from
the
forecasts,
and
the
forecasts
do
not
visually
appear
to
be
too
far
away
from
the
actual
observations
(when
compared
with
the
ARIMA
(1,1,0)
model).
However,
this
is
a
random
walk
model
which
is
not
particularly
useful
as
there
is
just
an
error
term
in
the
model
which
does
not
give
us
any
intuition
behind
the
model.
6
We
tested
other
models
including
more
AR
terms,
MA
terms
and
combinations
of
both
to
ensure
we
could
not
build
a
model
with
better
forecast
evaluation
statistics.
20
Cointegration
Model
To
be
able
to
create
a
Vector
Error
Correction
Model
(VECM),
all
series
need
to
have
the
same
level
of
integration;
in
our
case,
all
be
I(1).
From
our
initial
unit
root
tests
we
showed
that
all
the
variables
are
I(1)
as
they
are
stationary
in
their
first
difference,
therefore
we
can
use
them
to
create
a
VECM
as
long
as
they
show
a
cointegrating
relationship.
When
creating
a
VECM
we
are
looking
to
see
if
there
are
any
variables
which
restore
the
model
to
the
long
run
equilibrium,
if
there
is
cointegration
then
at
least
one
variable
needs
to
respond
to
this
long
run
equilibrium.
To
check
for
cointegrating
relationships
we
performed
two
tests;
the
Engle-
Granger
Test
and
the
Johansen
Test.
As
the
Engle-Granger
Test
doesnt
test
for
more
than
one
cointegrating
vector
it
will
not
be
as
powerful
as
the
Johansen
Test7.
Engle-Granger
Test
The
null
hypothesis
for
the
Engle-Granger
Test
is
that
the
series
is
not
cointegrated;
therefore
testing
that
the
residuals
from
the
model
have
a
unit
root.
To
be
able
to
say
that
there
is
a
cointegrating
relationship
between
two
variables,
in
addition
to
stating
that
they
are
all
I(1),
we
would
need
to
reject
this
null
hypothesis
of
non-cointegration.
7
As
we
have
7
variables
we
could
theoretically
have
6
cointegrating
relationships.
21
When
testing
for
a
unit
root
in
the
residual
we
obtain
a
p-value
of
0.02372,
which
is
lower
than
0.05,
therefore
at
the
5%
level
of
significance
we
can
reject
the
null
of
non-cointegration
(Figure
17).
This
result
indicates
that
there
is
a
cointegrating
relationship.
As
the
Engle-
Granger
Test
is
not
as
reliable
as
the
Johansen
Test
when
there
are
more
than
2
variables,
we
conducted
the
Johansen
to
ensure
our
result
holds.
Johansen
Test
This
is
a
more
credible
test
to
use,
as
we
have
more
than
two
variables
that
we
are
testing
for
a
cointegrating
relationship.
The
outcome
from
the
test
will
tell
us
how
many
cointegrating
vectors
we
have
within
our
variables.
With
all
other
tests
conducted
in
this
report
so
far
we
have
carefully
chosen
the
lag,
either
choosing
the
cube
or
fourth
root
of
the
sample
size.
However,
for
the
Johansen
Test,
we
gather
this
from
a
different
method
-
the
VAR
lag
selection.
22
Having
selected
the
relevant
lag,
we
can
now
use
this
in
the
Johansen
Test
to
discover
how
many
cointegrating
vectors
our
data
set
contains.
When
carrying
out
our
tests
we
started
by
using
a
restricted
trend
because
of
the
pattern
evident
in
the
time
series
plots8.
From
Figure
19
we
are
able
to
see
we
do
not
reject
the
null
of
one
cointegrating
vector,
as
0.2122
is
the
lowest
rank
to
be
significant
at
the
5%
level,
therefore
indicating
that
there
is
just
the
one
cointegrating
vector
in
our
data
set.
Additionally
the
Lmax
test,
as
well
as
the
alternative
Trace
test
(shown
above),
also
suggest
a
rank
of
1
as
their
values
are
larger
than
0.05.
As
it
is
not
clear
whether
a
restricted
or
unrestricted
trend
is
to
be
used,
we
continued
to
test
with
the
unrestricted
trend
to
ensure
we
ended
with
the
same
result.9
8
The
unrestricted
trend
was
also
tested
for
certainty.
9
To
further
guarantee
that
we
had
discovered
all
cointegrating
vectors,
we
repeated
the
Johansen
Test
twice
with
both
the
restricted
and
unrestricted
deterministics;
once
with
a
lag
of
3
and
then
with
a
lag
of
1.
Both
times
we
ended
with
the
same
inference
therefore
confirming
our
earlier
results.
23
Now
that
we
have
discovered
the
number
of
cointegrating
vectors
in
our
data
set
we
are
able
to
go
forward
and
create
a
VECM.
In
this
modelling
instance,
when
pertaining
how
many
lags
we
should
include
in
the
model,
we
use
an
iterative
process
whereby
we
check
for
autocorrelation
(AR)
at
0
lags,
continuing
until
we
have
a
lag
length
with
no
present
AR.
We
were
aware
that
when
using
GRETL
to
create
a
VECM
an
extra
lag
needed
to
be
added
to
the
number
entered,
as
in
GRETLs
input
section
a
lag
of
1
would
in
fact
mean
0
lags
in
GRETL.
For
example
when
we
were
testing
up
from
0
lags
to
see
if
there
was
any
AR,
we
entered
lag
1
and
rank
1.
When
choosing
the
deterministics
in
the
creation
of
this
model
we
used
both
restricted
and
unrestricted
trend,
our
results
held
the
same
inference
for
both.
When
using
0
lags
(therefore
1
lag
in
GRETL)
we
found
AR(1)
in
the
output
therefore
continued
testing,
but
now
with
a
higher
lag.
At
a
lag
of
1
(therefore
2
in
GRETL),
we
found
that
there
was
no
AR
(1)
or
AR
(2)
present
in
the
data
as
all
Ljung-Box
Q
p-values
are
greater
than
5%
(see
Figures
20
and
21).
Figure
20
Evidence
of
no
AR
(1)
24
VECM
Output
25
When
building
this
VECM
model
with
7
variables,
we
find
that
the
speed
of
adjustment
coefficient
(EC1)
for
the
Exxon
Mobil
and
US
unemployment
rate
variables
are
positive
with
a
coefficient
of
0.0752823
(Figure
24)
and
negative
with
a
coefficient
of
0.0101341
(Figure
27)
respectively.
This
follows
our
original
theoretical
predictions
regarding
the
positive
and
inverse
relationship.
Both
of
the
variables
are
significant
at
the
5%
level
with
a
p-value
of
0.0101
and
2.08e-11.
So
in
a
situation
out
of
the
equilibrium
Exxon
Mobil
and
the
US
unemployment
rate
will
help
restore
the
long
run
equilibrium.
All
the
other
variables
(Brent,
WTI,
DJIA,
US
10-year
bond,
GBP/USD)
are
weakly
exogenous10.
Since
we
were
able
to
conclude
that
the
series
are
cointegrated,
we
cannot
use
a
regression
equation
in
the
differences
because
the
OLS
will
omit
the
error
correcting
mechanism.
Therefore
we
will
forecast
using
this
VECM11.
10
If
the
variable
does
not
respond
to
the
disequilibrium,
it
is
weakly
exogenous.
11
We
tried
to
improve
this
model
by
adding
in
additional
variables
to
better
capture
the
cointegrating
relationship.
We
added
the
VIX
index
and
the
reciprocal
of
the
existing
US
unemployment
rate
as
variables.
However,
neither
of
these
improved
the
model
significantly,
so
they
are
excluded
from
the
model.
26
Period
Actual
Prediction
VECM
Standard/Error
2012:09
2012:10
2012:11
2012:12
2013:01
2013:02
2013:03
2013:04
2013:05
2013:06
2013:07
2013:08
112.86
111.71
109.06
109.49
112.96
116.05
108.47
102.25
102.56
102.92
107.93
111.28
118.18
119.82
120.74
121.62
122.49
123.36
124.19
125.01
125.81
126.59
127.36
128.11
5.217
9.084
12.264
14.894
17.161
19.188
21.043
22.767
24.388
25.922
27.383
28.782
95%/Interval
Lower/Bound Upper/Bound
107.95
128.4
102.02
137.63
96.7
144.78
92.42
150.81
88.86
156.13
85.75
160.96
82.95
165.44
80.39
169.63
78.01
173.61
75.78
177.39
73.69
181.03
71.7
184.52
27
Actual
Prediction
112.86
111.71
109.06
109.49
112.96
116.05
108.47
102.25
102.56
102.92
107.93
111.28
113.94
114.52
115.11
115.69
116.27
116.85
117.43
118.01
118.6
119.18
119.76
120.34
VECM
95%0Interval
Lower0Bound Upper0Bound
102.4
125.48
98.21
130.84
95.12
135.09
92.61
138.76
90.47
142.07
88.59
145.11
86.9
147.96
85.38
150.65
83.98
153.21
82.69
155.67
81.49
158.03
80.37
160.31
Actual
Prediction
112.86
111.71
109.06
109.49
112.96
116.05
108.47
102.25
102.56
102.92
107.93
111.28
118.18
119.82
120.74
121.62
122.49
123.36
124.19
125.01
125.81
126.59
127.36
128.11
95%0Interval
Lower0Bound Upper0Bound
107.95
128.4
102.02
137.63
96.7
144.78
92.42
150.81
88.86
156.13
85.75
160.96
82.95
165.44
80.39
169.63
78.01
173.61
75.78
177.39
73.69
181.03
71.7
184.52
As
can
be
seen
in
Figure
31
the
spread
between
the
ARIMA
and
VECM
forecast
widens
the
further
you
move
away
from
August
2012
our
last
actual
data
point
not
held
for
the
holdout
period.
This
can
clearly
be
seen
by
the
prediction
for
August
2013
where
the
actual
was
$111.28,
and
the
ARIMA
forecasted
a
value
of
$120.34
and
the
VECM
forecasted
$128.11.
The
comparison
of
the
two
forecasts
evaluation
statistics
is
shown
below
in
Figure
32.
Forecast%Evalution%Statistics
Mean%Error%%%%%%%%%%%%%%%%%%%%%%%
Mean%Squared%Error%%%%%%%%%%%%%%%%
Root%Mean%Squared%Error%%%%%%%%%%
Mean%Absolute%Error%%%%%%%%%%%%%%
Mean%Percentage%Error%%%%%%%%%
Mean%Absolute%Percentage%Error%%%
Theil's%U%%%%%%%%%%%%%%%%%%%%%%%%%
Bias%proportion,%UM%%%%%%%%%%%%
Regression%proportion,%UR%%%%%%
Disturbance%proportion,%UD%%%%%%
ARIMA%(0,1,0)
)8.18
97.333
9.8657
8.18
)7.7096
7.7096
2.8027
0.68745
0.16811
0.14444
VECM
)14.644
254.02
15.938
14.644
)13.669
13.669
4.4773
0.84422
0.10096
0.054823
As
previously
observed
in
the
comparison
of
the
forecasts
to
the
holdout
period
in
Figure
31,
this
is
backed
up
by
the
above
forecast
evaluation
statistics.
This
can
be
seen
as
the
ARIMA
has
28
a
lower
Theils
U
statistic
as
well
as
lower
other
error
statistics.
From
this
data
we
would
choose
to
use
the
ARIMA
(0,1,0)
as
our
forecasting
model,
as
it
has
shown
to
be
a
more
accurate
forecasting
model
over
the
holdout
period.
However,
we
are
dubious
about
using
this
model,
as
aforementioned,
it
is
a
random
walk
and
this
is
not
a
very
good
forecasting
model.
29
12
We
re-tested
for
unit
roots
and
stationarity
on
the
complete
data
set
(including
the
hold
out
period)
and
found
that
all
results
were
the
same.
The
levels
were
I(1)
and
the
first
differences
were
stationary.
30
Period
2013:09
2013:10
2013:11
2013:12
2014:01
2014:02
2014:03
2014:04
2014:05
2014:06
2014:07
2014:08
2014:09
2014:10
2014:11
2014:12
2015:01
2015:02
2015:03
2015:04
2015:05
2015:06
2015:07
2015:08
Prediction
111.806
112.332
112.859
113.385
113.911
114.437
114.963
115.49
116.016
116.542
117.068
117.594
118.121
118.647
119.173
119.699
120.225
120.752
121.278
121.804
122.33
122.856
123.383
123.909
ARIMA
95%
Interval
Lower
Bound Upper
Bound
100.524
123.088
96.3774
128.287
93.3178
132.399
90.8211
135.949
88.684
139.138
86.8024
142.072
85.1144
144.812
83.5796
147.4
82.1702
149.861
80.8656
152.218
79.6504
154.486
78.5128
156.676
77.4432
158.798
76.4339
160.86
75.4785
162.867
74.5717
164.827
73.709
166.742
72.8866
168.616
72.1012
170.454
71.3499
172.258
70.6301
174.03
69.9397
175.773
69.2766
177.488
68.6391
179.178
Prediction
109.511
109.046
109.565
110.278
110.992
111.694
112.382
113.055
113.716
114.364
115.001
115.628
116.245
116.854
117.455
118.049
118.636
119.218
119.794
120.366
120.933
121.495
122.055
122.611
VECM
95%
Interval
Lower
Bound Upper
Bound
99.4176
119.604
91.6159
126.477
86.1234
133.008
81.8308
138.725
78.2108
143.774
75.0371
148.35
72.1841
152.579
69.5733
156.538
67.1534
160.279
64.8893
163.839
62.7555
167.247
60.7332
170.523
58.8076
173.683
56.9674
176.74
55.2033
179.707
53.5077
182.59
51.8743
185.398
50.2978
188.138
48.7738
190.815
47.2984
193.433
45.8682
195.997
44.4802
198.511
43.1318
200.978
41.8207
203.4
Figure 36 Comparison of both the out of sample ARIMA (0,1,0) and VECM forecasts
31
The
out
of
sample
forecasts
both
expect
the
price
of
Brent
to
increase
over
the
next
two
years.
The
ARIMA
model
has
less
of
a
slope
than
the
VECM,
however
starts
at
a
higher
level.
The
ARIMA
and
VECM
models
both
forecast
an
increasing
price,
with
values
of
$123.90
and
$122.60
per
barrel
in
August
2015,
respectively.
Conclusion
In
conclusion,
we
have
created
an
ARIMA
and
a
VECM
model
in
order
to
forecast
the
price
of
Brent
crude
oil.
Through
comparing
these
models
over
our
holdout
period
from
August
2012
to
August
2013
we
found
that
the
ARIMA
(0,1,0)
was
a
more
accurate
model
based
on
forecast
evaluation
statistics
(see
figure
32).
Firstly,
due
to
the
volatile
nature
of
the
variables
selected,
the
ARIMA
model
in
hindsight
may
not
have
been
the
best
option
to
use
as
its
characteristics
lend
it
to
best
fit
a
linear
mean
trending
data
series.
Also,
ARIMA
models
tend
to
be
good
at
forecasting
short-term
changes,
rather
than
long-term.
From
the
outset,
we
were
mindful
that
our
ARIMA
forecast
would
potentially
not
be
too
accurate
as
we
planned
to
forecast
two
years
out
of
sample.
Further,
since
our
model
was
structured
as
an
ARIMA
(0,1,0),
this
shows
that
the
price
of
Brent
crude
oil
follows
a
random
walk
pattern.
After
reaching
the
conclusion
of
no
AR
or
MA
terms
it
made
it
clear
that
an
ARIMA
model
was
possibly
not
the
model
to
use,
as
ARIMA
(0,1,0)
models
are
basically
impossible
to
forecast;
since
they
are
dependent
on
the
price
of
the
previous
period
and
an
error
term.
Secondly,
as
aforementioned
our
variables
are
all
focused
upon
the
US.
If
we
were
to
widen
our
scope
and
take
into
account
additional
variables
from
outside
of
the
US
then
we
could
have
captured
more
than
just
purely
one
economys
demand.
If
we
had
access
to
supply
figures
from
the
Middle
East,
Norway
and
other
major
producers
then
we
feel
we
would
have
been
more
32
able
to
produce
an
accurate
forecast.
Furthermore,
we
may
not
have
selected
the
most
significant
variables
that
are
best
related
to
the
price
of
Brent
crude.
For
instance
we
could
have
included
other
variables
such
as
inflation,
indicators
of
Eastern
economies
(especially
China)
and
measures
of
supply
disruption.
For
example
when
OPEC
was
formed
or
during
the
Iraq
war
there
were
a
spikes
in
oil
prices.
Thirdly,
our
data
sample
ranges
from
January
2000
to
August
2013
which
includes
economic
downturns
including
the
burst
of
the
dotcom
bubble
in
the
early
2000s,
the
events
of
September
11th
and
the
financial
crisis
of
2007/08.
The
inclusion
of
dummy
variables
could
have
benefited
our
model
through
reducing
the
impact
of
the
aforementioned,
therefore
potentially
increasing
its
forecasting
ability.
From
the
outlook
we
were
sceptical
as
to
whether
we
were
going
to
be
able
to
produce
an
accurate
and
practical
model,
as
if
this
was
possible
people
in
the
financial
industry
would
currently
use
such
models
and
make
significant
profits.
33
Appendix
Appendix
1
1. WTI
Crude
oil
prices
(USD):
http://www.eia.gov/dnav/pet/pet_pri_spt_s1_d.htm
2. Exxon
Mobil
Corporation
(XOM):
http://finance.yahoo.com/
3. Dow
Jones
(USD):
http://finance.yahoo.com/
4. 10-year
US
bond
yield
(%):
http://www.ecb.europa.eu/home/html/index.en.html
5. US
unemployment
rate
(%
of
population):
http://data.bls.gov/pdq/SurveyOutputServlet
6. USD/GBP
exchange
rate:
http://research.stlouisfed.org/fred2/graph/?s[1][id]=EXUSUK
34
Appendix
2
Sample
of
the
data
output
for
unit
root
and
stationarity
tests
performed
on
the
levels.
ADF
Brent
Levels
KPSS
Brent
Levels
35
KPSS
US
10
Year
Levels
36
Appendix 3
Sample
of
the
data
output
for
unit
root
and
stationarity
tests
performed
on
the
first
differences.
ADF
DJIA
First
Differences
37
38
Appendix
4
ARIMA
(1,1,0)
Output
39
Bibliography
Associated
Press.
"Apple
loses
title
of
world's
most
valuable
company
to
Exxon."
Fox
News.
April
17,
2013.
http://www.foxnews.com/tech/2013/04/17/apple-loses-title-world-
most-valuable-company-to-exxon/
(accessed
November
20,
2013).
DeCarlo,
Scott.
"Global
2000:
China
Takes
The
Lead."
Forbes.
April
17,
2013.
http://www.forbes.com/global2000/
(accessed
November
20,
2013).
Durden,
Tyler.
"The
Most
Important
Number
In
The
Entire
U.S.
Economy."
Zero
Hedge.
August
1,
2013.
http://www.zerohedge.com/news/2013-08-01/most-important-number-
entire-us-economy
(accessed
November
23,
2013).
Lazzaro,
Joseph.
"Why
A
US
Debt
Default
Could
Push
Oil
Above
$130."
International
Business
Times.
September
28,
2013.
http://www.ibtimes.com/why-us-debt-default-could-push-
oil-above-130-1412344
(accessed
November
23,
2013).
Nelder,
Chris.
"Oil
demand
shift:
Asia
takes
over."
Smart
Planet.
March
6,
2012.
http://www.smartplanet.com/blog/the-energy-futurist/oil-demand-shift-asia-takes-
over/?tag=search-river
(accessed
November
17,
2013).
Nguyen,
Lananh.
"
London
Overtakes
New
York
as
Brent
Oil
Beats
WTI."
Bloomberg.
August
1,
2012.
http://www.bloomberg.com/news/2012-08-01/brent-beats-wti-oil-trading-as-
ice-eclipses-nymex.html
(accessed
November
21,
2013).
"Brent-WTI
Spread
Expands
to
Seven-Month
High
on
Libya."
Bloomberg.
November
13,
2013.
http://www.bloomberg.com/news/2013-11-13/brent-oil-gains-as-protests-in-libya-
disrupt-supply.html
(accessed
November
21,
2013).
"Rising
Oil
Demand."
xfuels.com.
2011.
http://www.xfuels.com/global-issues/rising-oil-
demand.php
(accessed
November
16,
2013).
Smith,
Grant.
"Brent
Poised
to
Depose
WTI
as
Most-Traded
Oil
Futures."
Bloomberg.
November
26,
2012.
http://www.bloomberg.com/news/2012-11-26/brent-poised-to-oust-wti-as-
most-traded-oil-futures.html
(accessed
November
20,
2013).
40
41