Time Series Analysis of Saudi Arabian Oil Output WORKING PAPER Hendrik Blommestein Abstract: I estimate a VAR system deemed to be representative of oil market reality to analyze the co-movement of Saudi Arabian oil production within the larger oil market environment. The dynamic nature of the industry as well as multi variable interdependence within the oil system is accounted for by VAR methodologies. The dynamic interrelations characterizing the estimated system are studied in terms of Granger causality statistics, impulse response functions and forecast error variance decompositions. Keywords: OPEC, Saudi Arabia, Oil Supply, Dynamic Response JEL Classification: Q41, Q43, Q47
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Time Series Analysis of Saudi Arabian Oil
Output
WORKING PAPER
Hendrik Blommestein
Abstract: I estimate a VAR system deemed to be representative of oil market reality to analyze the
co-movement of Saudi Arabian oil production within the larger oil market environment. The dynamic
nature of the industry as well as multi variable interdependence within the oil system is accounted
for by VAR methodologies. The dynamic interrelations characterizing the estimated system are
studied in terms of Granger causality statistics, impulse response functions and forecast error
variance decompositions.
Keywords: OPEC, Saudi Arabia, Oil Supply, Dynamic Response
JEL Classification: Q41, Q43, Q47
1
1. Introduction: Research Overview
The entire oil production space is generated by the operations of national oil
companies (NOCs), seven major international oil companies (IOCs) and
independent operators, known simply as independents. These entities control
roughly 80%, 7% and 13% of the worldโs proven-plus-probable reserves, respectively
(WEO 2013, pg. 421). Demarcating NOCs from the remainder of the production
space is the fact that these entities are at a minimum majority owned by their host
governments. A substantial fraction of these (in terms of reserves and production)
being completely state-owned and directed.
As a consequence of ownership differences, fully floated producers like Shell or
Exxon are likely to exhibit differences in the drivers of their production and
reinvestment decisions compared with partially floated companies like Statoil or
Rosneft.1 Likewise, a partially floated producer is likely to have different decision
drivers to those of their fully state owned counterparts. An exhaustive study of the
different behavioral drivers along this continuum is beyond the scope of this work.2
However, a description of fully state owned producers within the context of the
workings of the OPEC cartel and then more specifically Saudi Arabia as its lead
producer informs this research in important ways. The research objective being to
model Saudi supply in response to events in the global crude oil market place as
executed through its national oil company Saudi Aramco over the period 1994:Q1-
2013:Q2.
There are numerous approaches that can be taken to model Saudi supply. This
ranges from highly detailed fundamentals based modeling (including taking explicit
account of Saudi budgetary needs) to statistical models which abstract away large
amounts of industry detail. The approach taken here is to estimate a vector
autoregressive system of oil production split between Saudi Arabia and the
remainder of the production space, a representative demand proxy and the oil price
levelโ allowing us to observe Saudi supply movements in context of movements in
the crude oil market as a whole with a minimum of restriction.
The minimal restrictions characterizing VAR models generally โ which allows for
modeling multiple endogenous variables jointly โ leads us to observe the dynamic
relations of all variables within the system. Thus the research objective rather
1 Partial here still referring to entities meeting the NOC criterion of being majority owned by their host
government. 2 Tordo (2011) provides a thorough overview of some key differences.
2
forces observations to extend to empirical results of the model in the form of (1)
variance decomposition for all variables within the system. Particular attention
however is paid to Saudi Arabia.
The paper is structured as follows. The literature review is two pronged covering
sections 2 and 3. Each section motivates distinct aspects of the research design.
Section 2 provides background information on the OPEC cartel and the
international pricing system, with implications for the sample period of this study.
Section 3 discusses NOCs more broadly and the reasons for the selection of the
Saudi NOC particularly. Section 4 details some properties of the data introducing
also the econometric model with the discussion of empirical results contained in
section 5. Concluding remarks are offered in section 6. References and the appendix
can be found in sections 7 and 8, respectively.
2. The Evolution of Crude Oil Pricing and OPEC: Implications for Sample
Period
The OPEC pricing system, while not altogether collapsing on a fixed date, had
effectively transitioned to a market based pricing system by 1988. Within little
more than the space of one decade, OPEC had asserted its control over 51% of daily
oil production (1973), presided over two historical price shocks โ in the process
realizing unprecedented revenues โ before finally succumbing to structural changes
in the world oil market. This culminated in the collapse of OPEC market share to
28% in 1985, and with it the OPEC administered pricing system following within
several years.3
The consolidation of control by OPEC in the 1970s was a historical process
occurring on the back of waves of nationalizations and bids for equity participation
in OPEC member-country producing operations by OPEC member governments.
Prior to the 1970s the majority of OPEC oil had been under the complete control of
multinational IOCs. In fact, from the discovery of oil in the Middle East at the
beginning of the twentieth century until the early 1970s, OPEC member countries
played no role in the production or pricing of crude oil (Fattouh, 2006).4
The two OPEC induced shocks occurring in 1973 (Arab-Israel War) and then 1979
(Iranian Revolution) were responsible for a respective 7.8% and 8.9% drop in world
3 The figures in this paragraph are drawn from Fattouh (2011). 4 A thorough historical account of the nationalization process can be found in Terzian (1985).
3
oil production. The adverse consequences of these supply shocks on the U.S.
economy are well documented by Hamilton (2003). The 1973 shock resulted in a
0.6% drop in U.S. real GDP with the 1979 shock resulting in a 3.2% drop in U.S.
real GDP. Numerous studies have tested and rejected the hypothesis that the
relation between oil prices and output could simply be a statistical coincidence,
including Daniel (1997), Carruth (1998) and Hamilton (2003).
However, while output has been a relatively quick response variable in these
instances, the structural response of the world oil system on both the production
and consumption sides, has played out gradually, with price obviously lagging until
such time as the supply-demand balance changes. On the supply side, a lagged
response by non-OPEC producers to tap sources in the North Sea, Alaska, Mexico
and elsewhere occurred over the course of years. EIA data shows non-OPEC
producers increasing market share from 48% to 71% from 1975 to 1985, with this
supply largely originating from these new producing locations as well as the Soviet
Union. On the demand side, dampened consumption due to efficiency drives and
substitution away from oil also took years to take effect. These demand side effects
served to decrease oil consumption by 13% between 1979-1981, in the United
States, Europe and Japan.5 Starting in 1980, oil prices began a rapid decline, with
the drop-off witnessing the price of oil fall from an average of $78.2 per barrel in
1981 to an average of $26.8 in 1986 โ remaining within the (real) range of $20-$40
until 2004.6
The triggering of these lagged responses in production and consumption
demonstrates the misunderstanding the cartel may have had of the dynamic nature
of the oil market. Importantly for this study, leaving this phase (1970-1988) out of
our statistical sample (which covers 1994:Q1-2013:Q4) is done in the argument that
OPEC, with Saudi Arabia as its lead producer, is likely to have learned from this
experience (unprecedented in its history up to that point) and hence can be
hypothesized to incorporate knowledge of these possible response dynamics into any
future production policy choices.
Furthermore, the changes in the oil market meant that the administered pricing
regime made possible by an overwhelming level of control over supplies (as was the
case in the 1970s and earlier, first by a handful of major IOCs and later by OPEC)
was supplanted by a dynamic supply base spreading beyond the reserve base under
the control of the cartel. This had the consequence of inaugurating a market-related
pricing system. As Fattouh (2007) notes, โThe adoption of the current market-
5 Energy Information Administration (EIA) data. 6 EIA data.
4
related pricing system represented a new chapter in the history of oil price
determination since it resulted in the abandonment of the administered oil pricing
system that had dominated the oil market from the 1950s until the mid-1980s.โ
The emergence of suppliers outside of OPEC and the growth of new buyers
increased the prevalence of armโs-length deals forming a reference for market
determined spot oil prices or barrels priced at the margin.7 Reference crude oil spot
prices sprang up near the source of physically traded volumes including Brent
which has its physical base in the North Sea and is processed at the Sullom Voe
terminal in the Shetlands Island, UK. The Brent market assumes a central stage in
the current oil pricing system, on the basis of which 70% of internationally traded
oil is directly or indirectly priced, including all export cargoes to Europe from Saudi
Arabia (Fattouh 2007, 2011).8 The Brent crude price is hence used as the
benchmark price in this study. Thus the approximate date of collapse of the OPEC
pricing regime (roughly 1988), which was followed by a market-related pricing
regime based on a larger supply base serves as the second motivation for analyzing
a sample period spanning 1994:Q1-2013:Q4.9
3. NOCs in General and in Specific: the Unique Role of Saudi Aramco in
World Oil
The existence of NOCs lends teeth to the notion that oil is a strategic commodity.
Control of 80% of crude oil reserves and a comparable percentage of daily
production draws attention to the behavioral drivers of these nationalized
producing entities. Their behavior, though varied, is rooted in common features of
NOCs as Tordo (2011) notes: โNOCs differ on a number of very important variables,
including the level of competition in the market in which they operate, their
business profile along the value chain, and their degree of commercial orientation
and internationalization. One thus needs to be mindful of possible over-
generalizations. On the other hand, most NOCs share at least some core
characteristics: for example, they are usually tied to the โnational purposeโ and
serve political and economic goals other than maximizing the firmโs profits.โ The
7 Armโs-length deals are referenced by price reporting agencies to provide a price for a benchmark
crude oil. Today many more transaction layers are referenced by price reporting agencies including
the forward market. 8 Along with Saudi Arabia, Kuwait and Iran also rely on the Brent benchmark using the so called
Brent Weighted Average (BWAVE), which is the weighted average of all futures price quotations
that arise for a given contract during the trading day. 9 Dating the sample series from 1994:Q1 rather than 1988 is due to constraints in data availability
from the U.S. Department of Energy. See Appendix A.1 for further data discussion.
5
consequences of being tied to a โnational purposeโ for operations however are
themselves highly variegated and differ a great deal from country to country and
hence from NOC to NOC. Large producing individual country NOCs have been the
subject of numerous studies including Norway (Al Kassim, 2006), China (IEA,
2011), Mexico (Moodys, 2003) and Russia (Victor, 2008). Studies covering NOCs
generally include Stevens (2008) and Tordo (2011).
OPEC NOCs are in a class of their own within the NOC production block given their
cartel affiliation. They are hence often studied in context of cartel theory.10 These
studies span classic textbook cartel to two-block cartel (Hnyilicza, 1976), to
dominant firm (Salant, 1976), to clumsy cartel (Adelman, 1980), to residual firm
monopolist (Adelman, 1982) to bureaucratic cartel (Smith, 2005). The question is
not whether OPEC (still) restricts output, as this is evident in spare capacity
buildups allowed by relatively lower per-barrel marginal production costs, but the
reasons behind these restrictions.11 This distinction puts the onus on studying
OPEC behavior with regard to output choices as is done in the referenced cartel
literature.
However, there are material challenges to treating OPEC as a coherent decision
making unit as is done in the aforementioned studies. This stems from the fact that
countries within the cartel have at times militarily engaged one another (Iraq and
Iran 1981, Iraq and Kuwait 1990) and have collapsed production for reasons not in
line with cartel rationale but rather in relation to internal problems. This includes
Venezuela in 2002, Iraq in 2003, Libya in 2011, Iran in 2012, and Nigeria in 2013,
whose concerns in many cases persist. Whatโs more, traditional number 2 and 3
producers within OPEC (Iran and Iraq), are effectively outside of the cartel quota
system due to longstanding production complications (Iraq) or due to present
sanction (Iran), limiting their ability to be regarded as enforcing OPEC policy.
This study therefore opts to focus on Saudi Arabia in particular given the fact that
the country has never been under sanction nor has it experienced an internal-
instability related cut in production. Its sole political maneuver with respect to
output adjustment is outside of our time series (the 1973 Arab-Israel War) โ
allowing us to focus on other explanations for variable movement. OPEC countries
10 OPEC member countries are, in order of joining, Saudi Arabia (1960), Iraq (1960), Iran (1960),
Kuwait (1960), Venezuela (1960), Qatar (1961), Libya (1962), the United Arab Emirates (1967),
Algeria (1969), Nigeria (1971), Ecuador (1973), and Angola (2007). 11 See chart in Appendix A.4 for OPEC spare capacity since 2004 and estimated OPEC marginal cost
per barrel chart out to 2020 in Appendix A.4.
6
comparable to Saudi Arabia in these dimensions play markedly smaller roles in
world oil. These include Qatar and the UAE, countries shouldering only 2.06 mbd
and 3.21 mbd of total production burdens respectively.12
Saudi Arabia led world (and OPEC) production with an average of 11.72 mbd of oil
produced in 2012 with the cartels second largest producer, Iran, producing 4 mbd in
2011 (prior to sanction imposition; EIA, 2014). Moreover, spare capacity, the
primary instrument in OPEC policy has historically been held in vast majority by
Saudi Arabia (see chart in Appendix A.4). Saudi Arabia is the significant swing
producer within the cartel and therefore world oil. For the purposes of this paper,
the remainder of the production space (which includes private producers, other
NOCs and marginal OPEC NOCs) is treated as a separate aggregate producing
entity, responding in a manner hypothesized to be distinct from Saudi Arabia.13
4. VAR Approach: Joint Behavior of Selected Time Series
The former sections have motivated the 1994:Q1-2013:Q4 sample period selection as
well as the justification for singling out Saudi Arabia as a unique player in global oil
markets. To account for the dynamic character of responses in the oil markets we
introduce a vector autoregressive (VAR) model to study Saudi Arabia in context of
the broader oil market environment.
Treating variables specific to the Kingdom of Saudi Arabia (denoted KSA) and those
concerning the rest of the world (denoted ROW) jointly โ which allows for dynamic
adjustment and the role of expectations โ lends itself to a vector autoregressive
approach. This approach, like much of the VAR literature, builds on Simsโs (1980a,
1980b) criticism of the then prevailing econometric identification methodologies and
the alternative identification logic resting on solving a macroeconomic system with
active expectations formations.
Similar to Sims (1980b), we here introduce a four-variable dynamic system as a
reasonable approximation of oil market reality. The time series used in the model โ
12 Mbd denotes million barrels per day. World oil production in 2012 was roughly 90 mbd. EIA data
for 2012. 13 The tradeoff in this construction is that ROW production will contain some degree of OPEC (and
therefore Saudi, due to leadership) response. On the other hand, omitting the marginal producers
that largely track Saudi production will detract from analyses of variable response to ROW
production movements. For this research the gain associated with excluding marginal OPEC (non-
Saudi) production from ROW data is taken to be less than the gains associated with having complete
data on ROW production for modeling ROW interactions with variables in the larger system.
7
crude oil production split between KSA and ROW (denoted ๐๐ก๐พ๐๐ด ๐๐๐ ๐๐ก
๐ ๐๐), a proxy
for ROW demand movements which is the percentage change in G20 GDP(denoted
โ๐บ๐ท๐๐ก๐บ20, excluding Saudi Arabia) and the price of Brent crude oil (denoted ๐๐ก
๐ค๐๐๐๐) โ
are described and charted in appendix A.1 with statistical summaries provided in
appendix B.1.14 These variables are taken to be relevant drivers of Saudi production
adjustments.
By separating Saudi Arabia out from the rest of the world we place its relationship
with the broader oil system at the center of the study. The selection of quarterly
data is based on Saudi Arabiaโs demonstrated ability to swing production on a
quarter-to-quarter basis in response to information available on similar โ and
smaller โ time horizons which includes price, ROW production and GDP change. In
using a dynamic model, the gradual response of production and demand, as
discussed in section 2, and the separating out of Saudi Arabian production from
total production โ as motivated in section 3 โ are hence all addressed in this
dynamic framework which also takes account of interrelations.
4.1 Stability, Lag Length and Dynamics
Visual inspection of our production data (see Appendix A.1) suggests a linear trend
in ROW production adding some 20 million bpd of production to world supply
between 1994 and 2013. The non-stationarity of this series is reflected in failure to
reject the null hypothesis of a unit root via an Augmented Dickey-Fuller (ADF)
test.15 Saudi production behaves markedly different displaying far larger
percentage swings in production. This reflects Saudiโs role in world oil markets as a
swing producer actively targeting price. Nevertheless total additions of 3 million
bpd over the sample period fail to belie the presence of a unit root identified in the
series vis-ร -vis an ADF test.
The real price series demonstrates a positive trend, especially as concerns the
increase in the per-barrel price after the year 2000, reflecting structural change in
world oil markets. This structural change is hypothesized to occur as a consequence
of consistent increases in the marginal cost of barrels outside of OPEC and other
large conventional reserve regions, serving new demand growth (Hamilton, 2014).
An ADF test rejects the null hypothesis of no unit root in the series.
14 Note: The statistical summaries are provided for 1st differenced data for those series found to be
non-stationary. See subsection 4.1. 15 Complete descriptions of test results for each series are available in Appendix B.2.
8
The identification of the stationarity of the % change in G20 (minus KSA) data
series rules out the possibility of estimating a vector error correction model (VECM)
as our set of series are not integrated of the same order.16 Instead, all series except
the I(0) process are transformed by taking first differences before estimating the
VAR.
According to VAR lag order selection criteria and allowing for the possibility of up to
12 lags (12 quarters) we find the recommended optimal lag specification of 3
quarters according to sequential modified LR test statistic (each test at 5% level),
FPE (final prediction error) and the AIC (Akaike Information Criterion).17
Identification beyond restrictions on lag length includes restrictions on the entry of
contemporaneous interactions into the VAR structure to sort out causality,
16 Furthermore a Johansen cointegration test finds no cointegrating relationships among the data
(Johansen, 1988). Rejecting the use of a VECM for our time series is further based on Engle and
Granger (1987) and Sims, Stock and Watson (1990). 17 Test results can be found in Appendix B.3.
9
As a VAR can be considered to be the reduced form of a dynamic structural equation
(DSE) model, choosing ๐ดโ1 is equivalent to imposing a recursive structure on the
corresponding DSE model.18 Following Kilian (2008) a recursive identification
structure is ordered based on assumptions regarding how quickly the different
variables respond.19 The ordering of the variables in ๐๐ก (where also the matrix ๐ดโ1 is
argued to be lower triangular) is based on the assumption that benchmark prices
for oil, reflecting the value of the marginal barrel, is taken to be the most quickly
updated variable in the system being the outcome of trades covering spot and
forward markets. Saudi Arabia, in contrast to ROW producers, quickly responds to
the price environment maintaining swing production capacity for this purpose.20
ROW producers, while certainly responsive to the price environment, react as more
conventional business entities hence responding more slowly than the OPEC cartel-
leader who is actively targeting price. Finally, as is documented by Hamilton (1983,
2003), GDP responds relatively sluggishly to oil price changes, hence occupies the
last place in the recursive structure.
5. Empirical Results
The model estimates are summarized in the following three sub-sections with
Granger-causality tests reported in 5.1, impulse responses functions (IRFs) in 5.2
and forecast error variance decompositions (FEVDs) reported in 5.3. Following
Stock and Watson (2001), these statistics are taken to be more informative than the
estimated VAR regression coefficients or ๐ 2statistics given the complicated
dynamics in the VAR.
5.1 Granger-Causality Statistics
Causality as defined by Granger (1969) is dealt with in the context of the estimated
VAR(3). The idea being if series 1 affects series 2, the former should help improving
the predictions of the latter variable. Granger causality thus reflects predictive
causality (as opposed to โrealโ causality) given past values of the interrelated time
series. The null hypothesis in the following test results is that the lagged regressors
(three lags taken together) do not help predict the dependent variable in the
regression. 18 The centrality of the identification problem for sound empirical interpretation of estimation results
warrants a summary discussion of some of the mathematics of the Cholesky decomposition and its
relation to structural VARs which is hence provided in section D of the Appendix. 19 For brevity we will at times refer to the recursive identification structure imposed on the VAR as
the โWold causal orderingโ. See Appendix D for more information. 20 The response of Saudi swing production to price movement is further analyzed in Appendix A.3.