International Journal of Forecasting 7 (1991) 299-315 299

North-Holland

Using belief networks to forecast oil prices

Bruce Abramson
University of Southern California, Department of Computer Science and Social Science Research Institute,
Los Angeles, CA 90089-I I I I, USA

Anthony Finizza
ARCO, 515 South Flower Street, Los Angeles, CA 90071, USA

Abstract: Belief networks are knowledge-based models, developed by segments of the artificial intelligence
and decision analysis communities, that have the potential to become important forecasting tools. ARCOI,
currently under development at the Atlantic Richfield Company (ARCO) and the University of Southern
California (USC), is the most advanced implementation of these models in a financial forecasting setting.
ARCOI’S underlying belief network is an artificial intelligence knowledge base; it models all variables
believed to have an impact on the crude oil market. Decision analytic elicitation techniques collect
information about the market’s variables, their value ranges, and their interrelationsships. A pictorial
market model - developed on a MAC II - facilitates consensus among the members of the forecasting team.
The system forecasts crude oil prices via Monte Carlo analyses of the network. Several different models of
the oil market have been developed; the system’s ability to be updated quickly in light of recent events in
the Persian Gulf highlights its flexibility.

Keywords: Belief networks, Influence diagrams, Simulation, Artificial intelligence, Decision analysis,
Judgement, Energy forecasting, Oil markets.

1. Introduction ARCOI’S underlying model is a belief network, a

In many respects, modeling is the key to fore-
casting; forecasters attempt to model the past and
the present in a manner that allows them to infer
the future. Forecasters, of course, are far from
alone in their use of models. The 20th century has
seen modeling become a dominant theme in many
of the natural and social sciences [see, e.g. von
Neumann and Morgenstem (1944) or Simon
(1969)]. Differences among the sciences, however,
have led to divergent approaches, orientations,
and tools; exchanges among users of disparate
modeling techniques have been few and far be-
tween. The ARCOI knowledge-based forecasting
system bridges some of these gaps by addressing
concerns common to users of a variety of model-
ing tools.
graphical structure that maps relationships among
variables, in this case the economic and political
factors that affect the oil market. Recent research
on the applicability of belief networks to diagnos-
tic problems has led to some powerful medical
expert systems [see Andreassen, Woldbye, Falck,
and Andersen (1987) and Heckerman, Horvitz,
and Nathwani (1990)]. ARCO~ marks the first re-
ported application of these models to forecasting.
The leap from diagnostic systems to forecasting
systems, however, should not be difficult; the rela-
tionship between diagnosis and forecasting has
long been recognized and discussed [see, e.g.
Einhorn and Hogarth (1982)]. Our use of a belief
network model to forecast oil prices touches upon
topics generally discussed in artificial intelligence
(AI), decision analysis (DA), econometrics, psy-

0169-2070/91/$03.50 0 1991 - Elsevier Science Publishers B.V. All rights reserved

300 B. Ahrumson. A. Finizra / Umg belief networks to forecast odpnces

chology, statistics, political science, energy eco- Forecasting: Most combined forecasts (or fore-
nomics, and other fields. Many of these topics, casting techniques) begin with multiple (complete)
however, manifest themselves in non-standard forecasts and then apply either a linear combina-
ways. Points of particular interest include: tion rule or a judgmental adjustment [Clemen
Econometrics: In a very general sense, any (1989)]. ARCOI'Sunderlying network explicitly im-
model that starts with an economic theory posit- corporates at least three forecasting techniques -
ing a relationship among variables and then at- time series data, econometric models, and expert
tempts to determine the precise numeric character judgments - directly into the model. It does this
of that relationship is an econometric model. Un- by forcing the expert to provide all assessments
der this definition, ARCOI'Snetwork (both in its (whether judgmental or data driven) at the same
entirety and in many of its parts) certainly quali- time and to verify that only direct relationships
fies as such. Unlike most econometric models. are captured in the network. In other words,
however. belief networks split the model from its ARC0 facilitates combination at the assumption
solution mechanism. level rather than at the conclusion level. This em-
Political anabsis: The vast majority of political phasis on explicit assumptions also ensures that
analyses are stated qualitatively. As a result. judg- any scenario generated by the network is inter-
mental political variables are rarely incorporated nally consistent.
into formal models on equal terms with their This paper describes the modeling effort that
data-driven economic counterparts. Belief net- went into the ARCOI system, with a particular
works require an added amount of quantitative emphasis on issues that should concern forecas-
precision in the assessment of judgmental varia- ters. Since many readers may be unfamiliar with
bles and thus allow political variables to be incor- the origins, uses, and basic terminology of belief
porated as an integral part of a forecast, rather networks, Section 2 provides a brief introduction.
than as an afterthought. Section 3 highlights some of their potential appli-
Decision analysis: Decision analysts have long cations to forecasting, most of which are just
been concerned with the psychologically valid beginning to be studied and exploited. Section 4
elicitation, refinement, and quantification of judg- then describes ARCOI as a case study of belief
ments - in particular of judgments about uncer- networks for economic forecasting. A brief
tainty and value. The construction of a belief summary and a few conclusions are offered in
network to model a domain - such as the oil Section 5.
market - makes heavy use of decision analytic
tools. Since these networks are generally designed
as reusable domain models rather than as one-time 2. Belief networks
solutions to decision problems, however, they en-
able these tools to be used not only in decision Belief networks are a class of models that have
settings but also in other information processing recently become important to researchers at the
tasks, such as diagnosis and forecasting. intersection of AI and DA. Despite their underly-
Artificial Intelligence: AI popularized the idea ing sophistication, belief networks are conceptu-
of separating a model from its solution mecha- ally simple. Any directed acyclic graph in which
nism. Standard rule-based expert systems, for ex- (i) nodes represent individual variables, items,
ample, build large rule bases describing a wealth characteristics, or knowledge sources, (ii) arcs
of knowledge and then, at a later date, apply demonstrate influence among the nodes, and (iii)
case-specific data to answer specific questions. functions associated with the arcs indicate the
Systems based on belief networks maintain this nature of that influence, qualifies as a belief net-
split, but they rely on probability, statistics, and work. The oldest, simplest, and best understood
decision theory (rather than on logic-based symbol networks - decision trees - have been studied
manipulation) to provide the theoretical underpin- extensively in many settings, including financial
nings of their models. In addition, the belief net- forecasting [Hertz (1964)]. Other members of the
work paradigm has flourished within the commun- family have been labelled ‘path analysis’, ‘causal
ity of researchers interested in the management of models’, ‘influence diagrams’, ‘recursive models’,
uncertainty [Ng and Abramson (1990)]. and ‘Bayes networks;’ work has been done on
 B. Abramson, A. Fmzrn / Using belief neiworks to forecast oil prices 301

economics illustrates more than the flexibility of

the model; it can also be used to demonstrate the
differences between diagnosis and forecasting. In
a medical setting, the typical sequence of events
begins with the observation of a symptom, which
in our example would be coughing (C). The di-
Exhibit 1. Example of belief network agnostic task is then to recalculate the posterior
probabilities of the two diseases, pneumonia and
common cold (P(A/C) and P(B/C) respec-
each of these models. The current wave of re- tively). These posteriors can be calculated by a
search on graphical structures more intricate than simple application of Bayes’ rule. In an economic
decision trees, however, has just begun and many setting, on the other hand, the task is generally
of the most basic questions about their applicabil- one of forecasting; supply and demand may be
ity, use, and limitations have yet to be answered. observed and input into the function, f, to yield a
The majority of the literature on belief networks projected price.
has concentrated on either their theoretical aspects The simple example of Exhibit 1 masks a mod-
as mechanisms for modeling and tracking uncer- eling technique of tremendous power and sophisti-
tainty (see, e.g. Lauritzen and Spiegelhalter (1988), cation. Each object in the domain - and its rela-
Pearl (1988) Shachter (1986, 1988)], or on the tionships to the objects that influence it (and to
design of a specific system [see Andreassen those that it influences) - may be studied in
Woldbye, Falck, and Andersen (1987), and Heck- relative isolation and modeled in its most natural
erman, Horvitz, and Nathwani (1990)]; their power and elegant form. The only restriction is that each
as general representational mechanisms has only node must contain a method for generating a
recently come into focus [Abramson (1992)]. single value of every combination of influences. In
Consider the belief network of Exhibit 1 for other words, any fully specified mathematical or
two trivial, but illustrative, examples. In this net- probabilistic relation can be incorporated into the
work, A, B, and C each represent distinct objects, model.
while the arcs from A to C and from B to C In general, the construction of a belief network
indicate that the values of A and B each influence requires the identification and encoding of four
the value of C. In a medical domain, for example, types of information:
A could represent the disease pneumonia, B the Variables: These are the items that might have
common cold, and C the symptom coughing. The an impact on the domain. Each variable is repre-
arcs could then represent probabilities, as follows: sented by a node.
the A to C arc could indicate that pneumonia I/a&s: Each variable may take on a different
causes coughing with probability p (P(C/A) = p), set of possible values. Some may be binary, while
and the B to C arc that a cold causes coughing others may require a full range of values, Values
with probability q (P( B/A) = q). In an economic are generally entered into a separate menu and are
setting, A might represent supply, B demand, C not shown as part of the network’s graphical rep-
price, and the arcs an economietric formula de- resentation.
scribing price as a function of supply and demand Influence: Although all variables in a domain
(P =f( S, 0)). (Note that in each of these cases, are interdependent, it is crucial that only strong
A, B, and C are parts of a larger network. Any degrees of influence be captured and that there be
useful medical model must consider more than no loops. Directed arcs (arrows) are drawn to
two diseases and one symptom, and a reasonable show the direction of influence between each pair
economic model would probably require some no- of related variables.
tion of a time lag and formulas indicating S = f( P) Dependence: If A influences B, then, by defini-
and D =f( P) in addition to P =f(S, 0). The tion, the value taken on by A should tell us
purpose of these examples, however, was simply to something about the range of possible values that
illustrate some uses of belief networks and not to might be taken on by B. This information is
say anything particularly interesting about medi- generally given either as a formula or as a prob-
cine or economics.) This shift from medicine to ability distribution and is entered into a separate
302 B. Abramson, A. Finizzu / Using belief networks to Jorecast or1price.r

menu (hidden from the network’s graphical repre- have led to a variety of interesting algorithms for
sentation). Variables modeled without ingoing calculating conditional independence among sets
arcs, or rooted nodes, are dependent on nothing of variables [Pearl (1988)], propagating informa-
else in the model; they capture exogenous varia- tion through a network [Lauritzen and Spiegel-
bles. Their values are thus dependent on those halter (1988). and reducing networks to solve deci-
assumed by none of the model’s other variables, sion problems [Shachter (1986, 1988)]. Although
and they are specified as either fixed constants or these algorithms were all devised for networks that
as unconditional (prior) probability distributions. address diagnostic and/or decision problems, most
These four elements of knowledge acquisition of them sould also eventually become useful to
and model construction are central to the concept forecasters. General information propagation,
of a belief network; they must be performed in however, is beyond the scope of this paper. Accu-
any setting, regardless of the system’s ultimate rate models must be devised before propagating
objectives. Variables, value ranges, and influence information becomes worthwhile. Since ARCOI
can usually be specified with relative ease. The represents the first detailed belief network-based
specification of relevant probability distributions forecasting system, the first issues that had to be
and/or formulaic relationships (for dependence), considered in its development were those related
on the other hand, is frequently frustrating and to modeling. Thus, the bulk of this paper is dedi-
time-consuming. The number of probability as- cated to a discussion of the modeling tools and
sessments required can quickly become intracta- techniques used in the design of ARCOI. Informa-
ble. This bottle-neck contributed to the failure of tion propagation and algorithmic development.
early approaches towards probability-based sys- however, should not be far behind: they are the
tem design [Edwards (1962)]. Many early system focus of follow-up investigations already under-
designers found themselves caught between the way.
unrealistic assumption of universal independence
and the intractable assumption of universal inter-
dependence. Belief networks allow dependence to 3. Knowledge-based forecasting systems
be modeled if, and only if. the domain expert feels
that there is a present and direct influence. Tracta- The AI literature on knowledge-based systems
bility can be further improved by interpolating divides all such systems into two components: a
across value ranges (possible only if the values are knowledge base and an inference engine. The
numeric), repeating relationships over time (possi- knowledge base is the underlying static model of
ble whenever belief networks capture phenomena the domain, while the inference engine allows
generally modeled as time series), including alge- case-specific data to be entered and case-specific
braic/econometric formulas (frequently possible inferences to be drawn. Systems based on belief
in economic domains), and applying divide-and- networks retain this division. The network itself is
conquer principles to the domain (possible in many a static knowledge base, constructed using soft-
diagnostic settings). All told, then, belief networks ware intended for modeling. Implicit and/or
facilitate the development of tractable, realistic case-specific information is extracted by a ques-
domain models. tion-answering processing engine. (The term ‘in-
Belief networks were originally introduced as a ference engine’ reflects a bias of the AI commun-
middle ground between psychologically valid elici- ity for problems of inference and diagnosis. The
tation procedures and mathematically valid repre- more general term ‘processing engine’ implies the
sentations of uncertainty [Howard and Matheson performance of some information processing task,
(1984)]. As such, they began with an understand- such as an inference, a diagnosis, a forecast, etc.)
ing of the heuristics and biases that typically Research on the introduction of forecasting
plague experts [Kahnemann, Slavic, and Tversky techniques into processing engines for belief net-
(1982)], the DA elicitation techniques that help works is in the infancy. Even at this early stage of
overcome these biases [von Winterfeldt and Ed- development, however, several interesting points
wards (1986)], and the axioms of Bayesian prob- can be made:
ability theory [Savage (1954) and Edwards, Lin- * Linear forecasting models can all be incorpo-
dman, and Savage (1963)]. These basic principles rated into a belief network. Since the only restric-
 tion on a node is that it return a single value of assumption set helps to trace differing opinions to
any combination of inputs, linear regression mod- their sources. In addition, consistent assumptions
els, time series, and econometric relationships are help generate consistent scenarios,
all valid descriptors of influence and dependence. *The explication of assumptions also allows
Thus, a network comprised exclusively of such them to be changed when their validity ceases.
relationships would be nothing more than a Recent shifts in Persian Gulf politics, for example,
graphical representation of a system of linear necessitated a change in the political analyses built
equations. into any model. In the network underlying AKCoI,
*One characteristic that distinguishes network political considerations had already been isolated.
models from their more standard linear and ma- Thus, the ~zo~~fu~j~~ of these networks makes it
trix counterparts is that they are static; the model easy to rework part of the model without redoing
and its solution are intentionally separated. This it entirely.
split is quite common in AI settings; it emerges *Every node in a network corresponds to a
from the split between knowledge-bases and domain variable and each one is specified with its
processing engines. It is similarly common in other relevant variable range. Variables may be instanti-
types of simulation models, but stands in stark ated one at a time. A complete instantiation of the
contrast to most linear and matrix modeling tech- network constitutes a single scenario. Multiple
niques. Consider, for example, a linear regression scenarios can form the basis of a Monte Carlo
or a time series model, whose purpose and un- simulation. This Monte Carlo procedure formed
derlying solution mechanism are both obvious: its the basis of ARCOI'S processing engine.
purpose is to assign a value to the independent The work that we have done on ARCOI, to date,
variable, and its solution mechanism is specified is but the tip of the iceberg. Forecasts generated
by the equation combining independent variables by Monte Carlo analyses of belief networks repre-
to generate that value. A quick glance at a belief sent the most obvious and straightfo~ard appli-
network, on the other hand, reveals neither pur- cation of forecasting technology. Furhter research
pose nor solution. is underway to (i) extend the model’s time frame,
* Belief networks are an outgrowth of investiga- (ii) incorporate further judgmental variables into
tions conducted by researchers who subscribe to the model, and (iii) test the model in an ex ante
the Bayesian interpretation of probability. To a context. This last item, in particular, promises to
classical statistician, probabilities are a measure of have a dramatic impact on forecasting; tests of the
frequency of occurrence. Thus, only data-driven model on historical data promise to serve in a
variables can be used in classic models. Bayesians, manner analogous to historical testing of time
on the other hand, view probabilities as a measure series models. The inclusion of judgmental varia-
of belief; Bayesian models frequently combine bles in the network, however, makes standard ex
data-driven variables with those that capture sub- anre testing difficult; the algorithms developed for
jective beliefs. When applied to economic settings, diagnostic networks [Lauritzen and Spiegelhalter
this property allows probabilistic beliefs (such as (1988) and Pearl (1988)], on the other hand, may
those prevalent in political analyses) to be com- be useful.
bined with more standard data-driven variables.
This technique is in sharp contrast to the general
approach of judgmental corrections to mechanical 4. ARCOI
forecasts.
*The modeling exercises underlying the design 4.1.Basics
of a belief network force all contributing experts
to make their assumptions explicit and consistent. Oil market models can be classified into two
This type of exercise leads to increased con- broad categories: optimization models and capac-
sistency, both within an indi~dual contributor’s ity utilization target models [Energy Modeling
assessment of the market and among all of the Forum (1982)]. Optimization models are generally
contributors. Since forecasts and models are gen- based on either depletable resource theory, domi-
erally built by contributors with expertise in dif- nant firm cartel theory, or both. Marshalla and
ferent aspects of the domain, a universally shared Nesbitt (1986) for example, incorporated both
304 B. Ahrumson, A. Finrzzu / Ustng beliefnetwurks to foremst or1prtws

theories into their popular, fully dynamic optimi- 4.2. Variubies

zation model. As they accurately depicted, how-
ever, optimization models are more useful for This section details the variables captured in
longer term projections than they are for the short ARCOI'S 1990 networks (see Exhibit 2). The varia-
term. Our aim was to develop a short term model bles can be broken into seven categories and eight
that can use political and qualitative variables time periods. The time periods are indicated by
while preserving the robustness of subjective in- suffixes in the figures and subscripts in the text.
put. We chose to develop the model along the Variables labelled ‘.l’, ‘ .2’, ‘.3’, and ‘.4’ correspond
lines of the capacity utilization target school or, as to the first, second, third, and fourth quarters of
some have denoted it, the ud hoc models of pro- 1990, respectively. Variables labelled “.d’. ‘.c’, ‘.b’,
ducer/consumer behavior [Gately (1984)]. The and ‘.a’ are historical; they correspond to the first,
central determinant of prices in the capacity utili- second, third, and fourth quarters of 1989. respec-
zation framework is the relationship of calculated tively. The seven categories. in turn, are historical,
production to exogenously determined capacity, annual, tax, demand, supply, politics. and price.
the result being a measure of market tightness that One other important bit of notation was used in
will indicate price pressure. the discussions of functional dependence: negative
Our choice of the capacity utilization frame- suffixes indicate quarters ago (e.g. ‘ _ 4‘ means four
work as the basis of ARCOI'S model highlights the quarters ago).
importance of subjective political variables. Since
capacity is exogenously determined and short term 4.2. I. Historical r!ariclbles
crude oil demand is highly price inelastic and The following variables represent events that
almost completely specified by seasonal patterns, have already occured, their values being retrieved
short term price forecasts can be (more or less) from the appropriate references. The variables dis-
reduced to forecasts of production. Production cussed in this section are indicated in Exhibit 2
levels, in turn, are essentially set by the political with alphabetic suffixes (i.e. ‘.a’, ‘.b’, ‘.c’, or ‘.d’).
decisions of the governments of oil producing
countries. OPEC’s Persian Gulf members (Saudi Demand: Total demand during the four quarters
Arabia, Iran, Iraq, Kuwait, UAE, and Qatar) are of 1989.
particularly important, because they tend to be the
only producers with substantial slack capacity.
Level: Prevailing price level at the beginning of
Thus, the inclusion of political analyses and ad-
the quarter. Price levels are used to calculate the
justments to production calculations appears to be
effect of fuel switching on demand. They are
much more appropriate than politically motivated
specified as one of: (i) < $10. (ii) < $13, (iii) < $15,
judgmental adjustments to mechanically forecast
and (iv) > $15.
price calculations.
ARCOI’S 1990 base case, shown in Exhibit 2,
OPEC: The price of OPEC oil (Saudi basket}
contains a one year model of the oil market that
during the four quarters of 1989.
was designed in early 1990 using historical data
through the fourth quarter of 1989 and subjective
assessments provided between November 1989 and US Prod: US production during the four quarters
February 1990. This base case network contains of 1989.
about 140 equations, many with time lags and
some expressed as conditional probabilities. Sec- NO Prod: Non-OPEC, non-US production
tion 4.2 enumerates these variables and relation- throughout 1989. These figures were derived by
ships. It is important to stress, however, that the subtracting US production levels from total non-
system is more than simply a model. ARCO~ was OPEC numbers.
designed to facilitate scenario development and
simulation exercises. One such exercise is dis- Core Prod: 1989 production levels for core OPEC
cussed in Section 4.3; it concerns a scenario devel- countries. Throughout this paper, OPEC core re-
oped in late-August/early-September 1990 to re- fers to the six Persian Gulf members, Saudi Arabia,
flect the new political realities of the Persian Gulf. Iran, Iraq, UAE, Kuwait, and Qatar.
305
306 3. Abrumson, A. Finizzu / Using belie/nerworks to forecus: od prices

Supply: Total free world oil supply during the and GT.4 are specified with similar “priors”, for
fourth quarter of 1989. use when the tax was not passed in any previous
quarter.
NIC: Normal inventory change is not technically
an historical variable. It was specified, however, as 01 Fee: a binary YES/NO variable. The probabil-
an exogenous constant in precisely the same ity with which an oil import fee will be passed.
manner as were the historical variables. (Inventory The probabilities differ from quarter to quarter.
analysis is one of the areas in which the model is and are conditioned on whether or not the gas tax
in the greatest need of reconsideration and revi- was increased. In addition, YES is an absorbing
sion). state. Oil import fees are also conditioned on
increases in the gasoline tax: gas tax increases are
4.2.2. Annd variables assumed to make the imposition of an import fee
The following variables are defined for the en- less likely.
tire year of 1990. They are not expected to change
over the course of the year.
GT Impact: translates the increased gasoline tax
from dollars-per-gallon to dollars-per-barrel. The
NC Cap: Non-core OPEC capacity is assumed to
hold steady throughout the year. relationship between tax policy and price, as im-
plemented in the 1990 networks, appears to be
incorrect; revisions are currently underway.
NC Prod: Non-core OPEC production is similarly
assumed to hold steady; it is dependent only on
capacity. 4.2.4. Demand variables
The following variables are used to calculate
World Growth: World GDP growth is broken into total free world demand. Demand calculations (at
four components: lesser developed countries least in the developed world) are more straightfor-
(LDC), Western Europe (WE), US, and Japan. ward than supply calculations because there are
Coefficients relating the four components of the fewer phenomena that allow a small group of
world economy to the single world growth varia- decision makers to affect the market. The only
ble were calculated by a linear regression on his- demand-side peculiarity identified, in fact, was
torical data. fuel switching, a decision on the part of the
managers of dual-fired utility plants to switch to
4.2.3. Tax vurilrbtes oil use. Its impact is restricted to times when
The following variables relate to US tax policy. prices are maintained below $15 per barrel and
Two types of potentially relevant legislation are demand may range as high as 2 MMBD (million
envisioned: an oil import fee and an increase in barrels per day).
the federal gasoline tax. If an oil import fee is
imposed, its presumed effect would be to place an Level: Prevailing price at the beginning of the
$18 per barrel floor on imported oil prices. In- quarter. This variable is set at one of four levels
creases in the federal gasoline tax could range and is relevant only to the calculation of fuel
from $0.01 to $0.50 per gallon. The tax per gallon switching. The relevant levels are (i) < $10, (ii) 6
is multiplied by a factor of 42 to get the effective $13, (iii) < $15. and (r~) > $15.
price increase per barrel.

Duration: length of time over which the current

GT: a binary YES/NO variable. The probability
with which the gasoline tax will be increased dif-
fers from quarter to quarter. In addition, YES is an
absorbing state: once an increase has been passed,
it is assumed to remain in place. This absorption
introduces an influence which remains from one
quarter to the next. Thus, GT.l is a rooted varia-
ble, specified with a prior distribution. GT.2, GT.3,
level has prevailed. This variable is also used to
calculate the impact of fuel switching.
Fuel Switching: Amount of increased demand due
to the adoption of oif by utility plants with dual-
fired furnaces. This variable is defined over the
[0,21 MMBD range at 0.1 MMBD intervals. It is
 B. Abramson,A. Finizza / Using belief networks to forecast oil prices 307

dependent on price level and duration and given Core Production: Amount produced by the core
as conditional probability distributions. OPEC countries. The interesting point here is
actually the calculation of overproduction due to
Fuel Switching =f(Level, Duration).
political tensions (or lack thereof). Thus, core pro-
Demand: total world demand, by quarter. De- duction is defined as demand plus a “political
mand is taken to be a function of the prevailing hedge factor”.
price (that is, the OPEC price from the previous Core Production = Core Demand + Politics.
quarter), price and demand from a year earlier
Cap Ut: Defined as the percentage of core OPEC
(thereby capturing seasonality), world growth, and
capacity being used for production.
fuel switching. The functional specification for
demand was determined by an hoc combination of Capacity Utilization = Core Prod/C Cap.
regression techniques and scenario analysis.
Supply: Total supply of crude oil. This identity is
simply the sum of production from the four
Demand =f(Demand_,, World Growth,
sources.
OPEC,, OPEC _ ,, Fuel Switching).
Supply = US Prod + NO Prod + NC Prod
4.2.5. Supply variables + Core Prod.
The following variables are used to calculate DeltaY Core Prod: The one year change in pro-
total free world supply. duction by core OPEC countries.
DeltaY Core Prod = Core Prod - Core Prod ~ 4.
US Prod: US production levels are assumed to
decrease linearly with time. Each variable is de- DeltaQ Core Prod: The one quarter change in
pendent only on US production in the previous production by core OPEC countries.
quarter.
DeltaQ Core Prod = Core Prod - Core Prod ~, .

No Prod: Other non-OPEC (i.e. excluding the US) DeltaY Sweet: The one year change in production
production is assumed to be approximately sea- of light, sweet crude, defined as all non-OPEC
sonally constant. production.

C Cap: The physical production capacity of core DeltaY Sweet = (NO Prod + US Prod)
OPEC countries. It is assumed to increase steadily -(NO Prod-, + US Prod-,).
throughout the year (or, at the very least, not to
decrease). 4.2.6. Political variables
The following variables introduce subjective
Delta I: the change in inventory levels. Inventory measures of core OPEC politics into the model.
patterns are viewed as strictly seasonal; ranges are Two different aspects of the political situation are
specified by quarter. considered: general intra-gulf relations and the
degree of conflict arising from unhappiness with
0 Call: the effective demand for OPEC oil, or the market share. These two variables are then com-
call on OPEC. This is an identity, defined as total bined and mapped into production to yield a
demand (world demand plus or minus inventory “political hedge to production” factor.
changes), minus the oil supplied by other sources.
Intragulf: A measure of the general political ami-
0 Call = Demand + Delta I - US Prod cability among the core members of OPEC.
- NO Prod.
Market Share: Designed to measure satisfaction
Core Demand: The effective demand for oil from
with market share (and thus, implicitly, compli-
core OPEC countries. This is an identity, defined
ance with quotas).
as the amount demanded from OPEC minus the
amount supplied by non-core countries.
Politics: Converts subjective political judgments
Core Demand = OPEC Call - NC Prod. into OPEC overproduction.
30x B. Abramson. A. Finizza / Using belief networks to ,forccust oil priws

4.2.7. Price variables oil market. Value ranges and precise dependence
The following variables directly represent the (algebraic, econometric, and probabilistic) were
price of various grades of crude oil; they are the omitted from the discussion, as they are from
heart of the model. Since price is, after all, the Exhibit 2’s picture of the network, because they
most interesting issue in the system, as well as the are not central to the model of the domain. They
one most often targeted for a forecast, price varia- are, however, crucial if the model is to produce
bles are crucial. Two different prices are tracked: any useful results. This distinction is characteristic
OPEC price (Saudi basket) and WTI (West Texas of belief networks; network structure (i.e. nodes
Intermediate, the benchmark US crude) price. and arcs) describes the domain, while network
Formulas for calculating both price and sweet/ parameters (i.e. historical data, prior probabilities,
sour differential are quite detailed. and numeric relationships) allow specific ques-
tions to be answered. In other words, the model
OPEC: The price of OPEC oil. Price is given as a illustrated in Exhibit 2 and described in Section
regression-weighted function of several variables: 4.2 captures one year of the oil market. A fixed set
(i) prevailing supply, demand, and price (from the of parameters (such as those that we used in our
preceding quarter), (ii) core OPEC’s capacity utili- studies) captures a time frame (the year 1990).
zation. (iii) the one-quarter change in core OPEC Thus. recasting the model for 1991 should require
production, (iv) the one-year change in core OPEC nothing more than reviewing and updating the
production, and (v) normal inventory change. network’s parameters. (It would, in fact, be this
simple were the network structure completely
satisfactory. Several potentially hazy areas - nota-
OPEC =f(OPEC,, Cap Ut, DeltaQ Core Prod,
bly US tax policy, inventory behavior, and Gulf
DeltaY Core Prod. Demand_ ,, NIC ,,
politics - have already been detected and are
Supply 1).
currently under revision. Once a fully satisfactory
network structure has been derived, however, up-
Time: A variable introduced to measure the trend dates should be restricted to parameter changes.
of a steadily increasing shortage of sweet crude. Structural changes should be few and far between
Time = Time_, + 1. and should correspond to fundamental changes in
the market.)
SS Diff: The sweet/sour differential is the dif- The discussion of variables and influence, then,
ference in price between sweet and sour crude. It was intended to convey a broad understanding of
is measured by a complex formula, dependent on the oil market. Forecasts, on the other hand, re-
(i) the relative shortage of sweet crude, as mea- quire data. The basic model was used to create
sured by the Time variable, (ii) the one year two sets of scenarios for 1990: a base case and a
change in core OPEC production, (iii) the one constrained capacity case. The base case was de-
year change in production of sweet crude, and (iv) signed in early 1990, and covers all four quarters
the impact of a gasoline tax, if passed. of the year; it follows the network outlined above.
The constrained capacity case was designed in late
SS diff =f(Time, DeltaY Core Prod, August/early September 1990, when a fundamen-
DeltaY Sweet, GT Impact). tal market shift occurred. It assumes that there
will be an effective boycott of Iraqi and Kuwaiti
oil and that all other producers will produce at
WTI: The price of WTI oil, subject to the possible
maximum capacity. A revised network, shown in
imposition of an oil import fee.
Exhibit 5, was constructed. It accounted for his-
WTI =J(OPEC Price, SS Diff. 01 Fee). torical data through the end of the second quarter
and produced forecasts for the third and fourth
quarters of 1990. These cases were developed to
4.3. Scenarios and datu demonstrate the system’s flexibility, not its accu-
racy. Recall that the system’s processing power is
The previous section detailed the variables and still restricted to Monte Carlo analyses; this entire
influences captured by ARCOI'S model of the 1990 phase of development must be viewed more as a
Exhibit 5. Constrained capacity network
proof-of-concept than as a demonstration of Exhibit 4
Exogenous variables used in the 1990 base case.
power.
Region Minimum Maximum
4.3. I. Base case Word GDP LDC 2% 6%
The base case for 1990 is described by the growth ’ us -1% 5%
network of Exhibit 2. Specific values for historical WE 0% 5%
Japan 1.5% 7%
and exogenous variables are given in Exhibits 3
and 4, respectively. Many of the probabilistic as- Quarter Pr[NO] Pr[ws] h
sessments and regression weights are currently Gas tax IQ 1989 0.97 0.03
under review and have not been released. Qualita- increase 2Q 1989 0.9 0.1
3Q 1989 0.X 0.2
tive analyses of US tax policy, fuel switching, and
4Q 1989 0.6 0.4
Gulf politics, however, certainly warrant further
discussion. Quarter Minimum Maximum
Delta 1Q 1989 - 3.0 MMBD - 1.5 MMBD
US tax policy is one instance of an important
inventory “ 2Q 1989 0.7 MMBD 1.5 MMBD
set of judgmental variables that does not fit into 3Q 1989 0.5 MMBD 1 .O MMBD
data-driven models. Since the US is the world’s 4Q 1989 ~ 0.5 MMBD 0.5 MMBD
largest consumer of oil, as well as its largest im-
Quarter Minimum Maximum
porter and one of its largest producers, US poli- Core OPEC IQ 1989 18.7 MMBD 19.1 MMBD
cies can affect the market in several ways. First, capacity ’ 2Q 1989 18.9 MMBD 19.6 MMBD
increased taxation could lead to slowly declining 3Q 1989 19.0 MMBD 21.1 MMBD
demand. Second. US taxes could have a direct 4Q 1989 19.1 MMBD 21.6 MMBD

Non-core OPEC 8.1 MMBD throughout 1990 ”

” Uniform distributions applied across each range.

Exhibit 3 h YES is an absorbing state.
Historical variables used in the 1990 base case. Normal inven- ’ Assumed nondecreasing. Otherwise. uniform distributions
tory change is included here for reasons of convenience. used.
Variable Quarter Value

Demand 1Q 1989 53.1 MMBD

2Q 1989 50.1 MMBD impact on the price of imported oil and conse-
3Q 1989 51.0 MMBD quently an indirect impact on world prices and on
4Q 1989 54.7 MMBD
domestically produced crude. Since crude oil spot
US IQ 1989 7.7X MMBD prices are typically quoted for WTI (as traded on
production 2Q 1989 7.74 MMBD NYMEX, the New York Mercantile Exchange),
3Q 1989 7.54 MMBD
the price of domestic oil is of central importance.
4Q 1989 7.40 MMBD
The analysis of US tax policy, as described in
Other (non-US) IQ 1989 22.72 MMBD Section 4.2.3, considered the possibility of two
non-OPEC 2Q 1989 22.36 MMBD
relevant taxes: an increase in the federal gasoline
production 34 1989 22.66 MMBD
4Q 1989 23.00 MMBD tax, and the imposition of an oil import fee (at an
$18 floor). The genera1 assessment was that the
Core OPEC 1Q 1989 12.9 MMBD
increased gasoline tax was the more likely of the
production 24 1989 13.9 MMBD
3Q 1989 14.9 MMBD two and that it would, if passed, decrease the
44 1989 16.1 MMBD likelihood of an oil import fee. (Note that the
federal gasoline tax was, in fact, increased for
OPEC IQ 1989 $15
price 2Q 1989 $17 1991 as part of the budget package passed in
34 1989 $16 October 1990.) Quarter-by-quarter assessments for
44 1989 $17 the probability of an increased gasoline tax are
Normal IQ 1989 2.3 MMBD given in Exhibit 4. The size of the increase, given
inventory 24 1989 1.3 MMBD that one was passed, is captured by the variable
change 3Q 1989 1.25 MMBD “GT Impact”. Several key probabilities were di-
4Q 1989 - 1.25 MMBD rectly assessed for this variable. Other probabili-
MMBD stands for million barrels per day. ties were calculated by interpolating across the
 B. Abramson, A. Fmizra / Usmg belief networks to forecast oil prices 311

range. The directly assessed values (expressed as two relevant characteristics of Gulf politics: (i)
part of a cumulative distribution function) were general political amicability among the core OPEC
Pr[Increase < $0.051 = 0.1, members, and (ii) each country’s satisfaction with
Pr[Increase < $0.071 = 0.25, its market share (and thus, implicitly, compliance
Pr[Increase < $0.171 = 0.75, with OPEC quotas). The third step placed these
Pr[Increase < $0.201 = 0.9, variables on subjective five-point scales (harmony-
Pr[Increase < $0.251 = 0.95, to-war and strict compliance-to-rampant cheating,
Pr[Increase < $0.501 = 1.0. respectively). The fourth step specified conditional
All prices are quoted in dollars per gallon. To probabilities relating the two, and the fifth step
convert to dollars per barrel, multiply by 42. Oil mapped them into oil production above or below
import fees were considered to be relatively un- natural demand. This type of political analysis is
likely, the most likely scenario that included a fee inexact and will certianly need to be refined, up-
requiring that the gasoline tax not be increased dated, and changed. (One such change is discussed
throughout the entire year. Even in that case, in the next section.) Its inclusion in the model,
however, the probability of an import fee being however, establishes a clear “political module”
passed in the fourth quarter was only 0.133. into which all updates can easily be inserted.
Fuel switching is another area that generally
eludes data-driven models. The ‘switching’ in fuel 4.3.2. Constrained capacity case
switching refers to a managerial decision at dual- The constrained capacity (or boycott) case was
fired utility plants to burn oil rather than natural designed after Iraq’s invasion of Kuwait on August
gas. This decision, unlike most others affecting the 2, 1990, The behavior of the oil market following
market, is an essentially macro-level demand deci- the Iraqi invasion and subsequent world reaction
sion; universal switching to oil could increase de- indicated that a fundamental change had occurred
mand by as much as 2 MMBD. ARCOI'Smodel and that all existing short-term models were in
included a fairly detailed analysis of fuel switch- need of (at least some) revision. The modularity of
ing. Without getting into specifics. the thrust of ARCOI'Sunderlying network facilitated these
the analysis is that oil warrants consideration at changes; the constrained capacity network is
prices below $15 per barrel. If low prices are shown in Exhibit 5.
maintained for an extended period of time, many The network that corresponds to this case in-
managers will opt for oil. The lower the prices and corporated several assumptions that actually made
the longer they are maintained, the greater the the analysis easier than it was in the base case.
demand. Twelve different curves were specified to First, and most obviously, it was designed seven
describe industry behavior under the assumptions months later and with two additional quarters of
of price levels held at or below (i) $10, (ii) $13, historical data. Thus, nodes corresponding to first
and (iii) $15 per barrel, for one, two, three, and and second quarter 1989 were dropped, and actual
four quarters. The variables “level”, “duration”, numbers for first and second quarters 1990 (see
and “ fuel switching” implemented this analysis Exhibit 6) were included. Second, it began with
and its impact on demand. the assumptions that the boycott of Iraq and
Political analyses and projections are neces- Kuwait would be effective and that everyone else
sarily softer (i.e. more subjective) than their eco- in the world would raise their production levels to
nomic counterparts. As a result, they are invaria- their maximum physical capacity. Since all non-
bly omitted from technical models and relegated OPEC producers were already assumed to be pro-
to the role of a posterior? judgmental adjustments. ducing to capacity and the production levels of
ARCOI'Sstructure helped initiate a quantitative - non-core OPEC contries has been fixed at 90% of
albeit subjective - analysis of core OPEC (Persian their 8.1 MMBD capacity, few changes were
Gulf) politics. The first step in the analysis lay in needed outside OPEC’s core. Core production,
realizing that politics is really significant only in however, was pushed up to 13.6 MMBD, the
its effect on production. Thus, rather than being combined capacities of Saudi Arabia, Iran, UAE,
an adjustment to price, politics is viewed as an and Qatar. Capacity utilization, originally intro-
adjustment to OPEC production. The second step duced as a measure of pressure on production and
of the analysis used this observation to identify the key to the model, was fixed at 1.0. As a result,
312 B. Ahramson. A. Fmizra / lJ.w~gbelief networks to Jorecast oil prices

Exhibit 6 that forecasting models attempt to capture our

Historical values of relevant variables from the first two
knowledge of the past and the present in a manner
quarters of 1990 (constrained capacity case).
amenable to the drawing of inferences about the
Variable Quarter Value future. ARCOI’S underlying belief networks cap-
Demand 1Q 1990 53.4 MMBD tured information about direct interrelationships
2Q 1990 51.1 MMBD among the variables affecting the oil market. Im-
Other non-OPEC IQ 1990 22.61 MMBD plicit in these direct relationships lies information
production 2Q 1990 23.42 MMBD about the market’s indirect relationships. The task
of the forecasting/processing engine must be to
OPEC price 1Q 1990 $17
24 1990 $14 explicate the indirect relationships between varia-
bles and future prices.
US production IQ 1990 1.46 MMBD
Although a wide range of statistical procedures
2Q 1990 7.22 MMBD
are (theoretically) available to ARCO~, only one
Core OPEC 1Q 1990 16.11 MMBD
simple technique has been fully implemented to
production 24 1990 16.31 MMBD
date: Monte Carlo analysis. The implementation
of Monte Carlo on the network was fairly
straightforward. Exogenous variables (represented
by rooted nodes, or nodes without ingoing arcs)
the entire political analysis module was dropped were specified as either constant values or as un-
from the network; despite the obvious volatility of conditional (prior) probability distributions. In
the political situation, it was assumed that the either case, assigning a single value to an exoge-
impact on production would be steady throughout nous variable was straightforward. Once all rooted
the rest of the year. (The model assumed that no nodes were instantiated, nodes pointed to only by
settlement negotiated before the end of the year rooted nodes (i.e. variables directly dependent only
would restore the situation ante). Fuel switching on exogenous variables) could similarly be instan-
was dropped from the analysis because the possi- tiated. This procedure continued until the entire
bility of sustained prices under $15 per barrel network (or. alternatively, the mid-network node
disappeared, and US tax policy was excluded (per- selected as the forecast’s target) was instantiated.
haps unreasonably) because it did not appear likely This assignment of a single value to each variable
to have much of an impact before the end of 1990. constitutes a single fully-specified scenario (i.e. all
Most of the other analyses remained as they had variables are instantiated). Multiple fully-specified
been in the base case. scenarios lead to a distribution of values across
One point worth noting is that this constrained the target variable and thus a probabilistically
capacity case is significantly outside the range of reported forecast. This procedure is a trivial gener-
possibilities that were envisioned when the base alization of the Monte Carlo studies performed on
case was designed. The initial political assessment, decision trees and value trees; it is guaranteed to
in fact, assumed that the pattern set during the work on belief networks because they are not
Iran/Iraq war would continue and that lack of allowed to contain any loops.
cohesion among Gulf countries would lead to Results were generated by Monte Carlo analyses
overproduction and low prices. The possibility of of the 1990 networks. They are not meant to be
a consumer boycott was not even considered. Nev- either complete or conclusive, but simply illustra-
ertheless, the model was flexible enough to be tive of the claim that the system works. The varia-
updated (quickly and painlessly) in the presence bles targeted by these forecasts were the network’s
of new data. sinks, namely WTI or WTIp (by quarter). The
WTIp variables recognize the possibility of an oil
4.4. Forecasts import fee placing an $18 floor on domestic oil
prices. They adopt the price calculated for WTI if
The first phase of ARCOl’s development stressed no fee is imposed, but report a price of $18 if the
model contruction. The system’s ultimate objec- fee is passed and the calculated price is less than
tive, however, is to use these models to forecast or equal to $18. Three sets of simulations were
the market. This paper opened with the assertion run. The first set studied the full base case, simu-
 B. Abramson, A. Finizza / Using belief networks to forecast oil prices 313

Exhibit 7 Exhibit 8
Means and standard deviations of the forecasts generated by a Means and standard deviations of the forecasts generated by a
Monte Carlo analysis of the original 1990 base case. All Monte Carlo analysis of the base case, updated using informa-
numbers in these table are approximate and quoted in dollars tion available in July 1990. All numbers in this table are
per barrel. approximate and quoted in dollars per barrel.

Quarter u 0 Quarter LL 0
1Q 1990 20.87 2.9 34 1990 19.18 2.5
2Q 1990 20.62 3.3 44 1990 20.79 4.4
3Q 1990 21.23 4.1
44 1990 21.84 4.4

standard deviation) from the forecast mean. In the

second quarter, average WTI price was $17.76,
lating the network shown in Exhibit 2. The second about $3 (or one standard deviation) from the
set retained the base case assumptions, but up- forecast mean. Thus, ARCOI'S forecasts prior to
dated the network with actual data for the first the Iraqi invasion of Kuwait (and the ensuing
two quarters of 1990 (as shown in Exhibit 6). The fundamental shift in the market) were relatively
third set simulated the constrained capacity case, accurate. The insertion of first and second quarter
using the network of Exhibit 5. (As the data will data, however, allowed us to re-run the simula-
show, however, this third set of simulations was tions for the third and fourth quarters. These
not really necessary.) In each of these simulations, updates, shown in Exhibit 8, were produced using
100 scenarios were generated for each target varia- data available in July 1990. They are well within
ble. (Simulations of 100 scenarios are not really the range of projections made by most of the oil
adequate. The small size was necessitated by im- industry’s analysts at the time. Unfortunately, the
plementation inefficiencies. Many of them have market has shifted sharply since then. When Iraq
already been corrected; our current implementa- invaded Kuwait, the US successfully led the
tion is running ten to twenty times as quickly.) United Nations to establish and effective embargo
The results of these simulations are shown in of Iraqi and Kuwaiti oil and, since all other pro-
Exhibits 7-9. ducers decided to increase production, the global
The base case forecast, shown in Exhibit 7 supply picture was altered drastically. Simulations
indicated a relatively flat market. All four quarters of our constrained capacity case yielded the very
generated average prices between $20 and $22, tight forecasts shown in Exhibit 9.
with an annual average of $21.14. The [$18,$21] The conditions underlying the constrained
range accounted for 246 of the 400 scenarios capacity scenario are sufficiently restrictive to re-
(61.5%) with just over half of them (50.5%) falling move virtually all uncertainty from the system;
between $19 and $21. The distribution of the detailed simulation and statistical analyses were
remaining 154 scenarios, however, was far from unnecessary. Under the system’s assumptions,
uniform. Only 34 scenarios (8.5%) projected prices supply is entirely fixed and demand is assumed to
at or below $17, and 23 of these actually hit the vary more or less in line with world GDP growth.
$17 level; the [$14,$16] range accounted for only Thus, prices generated under this scenario (at least
11 out of 400 possible cases (2.75%). Thus, the for the near term) are effectively fixed. Despite the
probability of a significant downward trend under volatility of spot prices throughout the third and
the base case conditions was highly unlikely. On fourth quarters of 1990, ARCOl’s constrained
the upside, however, there appeared to be more capacity forecasts were remarkably accurate; the
room for runaway prices. 67 scenarios (16.75%)
generated prices in the [$22,$25] range, 43 in the
Exhibit 9
[$26,$30] range (10.75%), and 10 in the [$31,$40]
Prices forecast by the constrained capacity case. All numbers
range (2.5%). in this table are approximate and quoted in dollars per barrel.
Recall that these results were based on assump-
Quarter Prices
tions available at the beginning of 1990. The ac-
tual average WTI price for the first quarter of 34 1990 25
44 1990 29-31
1990 was $21.70, within $1 (or about l/3 of a
314 B. Abramson, A. Finmu / Usmg beliefnetworks toforecast oil prices

(true) average prices were $26.31 for the third have yet to be explored. The underlying software
quarter and $31.91 for the fourth. Technical vola- is evolutionary; it grows in response to need. Since
tility, however, does highlight a potential problem the first crucial stage of the system’s development
facing the system. The networks discussed in this was the construction of a belief network model of
paper all focus on market fundamentals. Volatility the domain, work to date has emphasized model-
caused by war fears, unusually high risk factors, ing rather than forecasting. As a result, this paper
and other technical factors tends to elude funda- was intended more as a proof-of-concept than as a
mental analyses. In a disequilibrated (or day- demonstration-of-power. It was designed to intro-
traded) market, forecasts produced by ARCO~ are duce belief network models to the forecasting
unlikely to be useful. In a stable, fundamental- community through a detailed and relevant real-
based setting, however, the information captured world case study.
by ARCOI'S network does appear to model our New approaches to the design of a knowledge
understanding of the crude oil market in a manner intensive model or a knowledge-based system, such
amenable to producing relatively accurate fore- as those underlying ARCO~, stand at the cutting
casts. edge of several different fields. As the needs and
ARCO~‘S base case 1990 forecast was, in fact, abilities of these fields come into clearer focus, the
surprisingly good. Although the $20 prices preva- system should continue to evolve. It is impossible
lent in late 1989 and early 1990 were higher than to pinpoint ARCOI’S ultimate completion date or
they had been for most of the previous few years, its ultimate set of capabilities; evolutionary
ARCOI allocated a relatively large probability to a processes tend to defy descriptions of their end-
runaway upside (a 30% chance of prices between points. Nevertheless, ARCOI, at its current stage of
$22 and $40 per barrel) and a much smaller prob- development, should interest a researchers in many
ability to a runaway downside (only an 8.5% fields. The base case model described in this re-
chance of prices at or below $17 per barrel). A port is a carefully constructed and highly
qualitative description of this forecast and its im- sophisticated model of the 1990 oil market. It is a
plicit recommendation, then, would be: “Assume graphical equivalent of around 140 equations,
that the market will remain in the high teens and many of which include time lags and some of
low twenties, but prepare a contingency plan for which are specified as conditional probability dis-
higher prices. There are many low-probability/ tributions rather than as linear models. It is also
high-impact scenarios that could lead to high designed in a manner that facilitates scenario gen-
prices, but relatively few that will lead to lower eration and simulation exercises. The model repre-
ones.” Given the market’s performance during sents a high degree of convergence of concerns. It
1990, this type of forecast was about as accurate is, however, far from complete. The network will
as possible. The system, then, is off to a rather undoubtedly require fine-tuning, the ease with
promising start. Whether or not this performance which it can be extended to other years or longer
can be continued, however, remains to be seen. term models needing to be considered, and the
relationship between the model’s assumptions
(captured in the network) and its conclusions
(generated by simulation) requiring study. Several
5. Summary potential areas for major improvements in the
software have already been targeted. Evaluations
ARCO~ is a knowledge-based system designed and analyses need to be conducted. All of these
to help the members of ARCO’s corporate plan- issues are currently being considered; the system’s
ning group who are involved with forecasting the evolution is continuing, and it it spawning a series
price of crude oil. The system is based on a belief of related research projects. Some sources have
network, a type of graphical model that is rapidly claimed that network structures promise to have a
gaining popularity in both the AI and DA re- revolutionary impact on forecasting and planning,
search communities. ARCOI’S construction was in- akin to the impact of linear and matrix models in
volved and time-consuming. As the first reported the early 1970’s [The Economist (1989)]. Although
forecasting system of its type, it suggested many these claims are premature, the potential does
interesting basic research issues, most of which seem to be present. The material presented in this
 B. Ahrumson, A. Finizza / Using behef networks to forecast od prrces 315

paper represents a first, small step towards realiz- Howard, R.A. and J.E. Matheson, “Influence Diagrams”, in:
R.A. Howard and J.E. Matheson. eds.. 1984, Readings on
ing that potential.
the Principles and Applications of Declsion Analysis, Vol. II
(Strategic Decisions Group, Palo Alto, CA) 721-762.
Kahneman. D., P. Slavic and A. Tversky, eds.. 1982. Judge-
Acknowledgements ment Under Uncertainty: Heuristics and Biases (Cambridge
University Press, Cambridge, England).
Lauritzen, S.L. and D.J. Spiegelhalter, 1988, “Local computa-
ARCOI’S greatest strength lies in the team as-
tions with probabilities on graphical structures and their
sembled to make it a reality. Help was provided applications to expert systems”, Journal of the Royal Statis-
by Mikkal Herberg, Peter Jaquette, and Paul Tos- tical Societ_v Series B 50(2), 157-224.
setti, all of ARCO’s corporate planning group. Marshalla. R.A. and M. Nesbitt, 1986, “Future world oil prices
The code underlying the system was written (for a and production levels: An economic analysis”. The Energv
Journal 7(l), l-22.
MAC II) by Keung-Chi Ng, a recent graduate of
Ng. K.C. and B. Abramson, 1990. “Uncertainty management
USC’s doctoral program in computer science. This in expert systems”. IEEE Expert 5(2), 29-48.
work was supported in part by the National Sci- Pearl, J., 1988. Probabilistic Reasoning m Intelligent svstems
ence Foundation under grant IRI-8910173 to (Morgan Kaufmann, San Mateo, CA).
Bruce Abramson. Savage, L.J., 1954, The Foundations of Statistics (John Wiley &
Sons, New York).
Shachter, R.D., 1986, “Evaluating influence diagrams”. Oper-
ations Research, 34(6). 871-882.
References Shachter. R.D.. 1988, “Probabilistic inference and influence
diagrams”, Operations Research. 36. 589-604.
Abramson. B.. 1992, “On knowledge representation in belief Simon, H.A., 1969. The Sciences of the Artrfwal (MIT Press,
networks”, in: R. Yager, B. Bouchon-Meunier and L. Zadeh. Cambridge, MA).
eds., Information Processing and the Management of Uncer- van Neumann, J. and 0. Morgenstem, 1944, Theon, of Games
taint-v (Springer-Verlag). Berlin-New York) (forthcoming). and Economic Behavior (Princeton University Press, Prince-
Andreassen. S., M. Woldbye, B. Falck and S.K. Andersen, ton, NJ).
1987, “MUNIN - A causal probabilistic network for the von Winterfeldt. D. and W. Edwards, 1986, Decision Ana!vsis
interpretation of electromyographic findings”, Proceedings and Behavioral Research (Cambridge University Press,
of the 10th International Joint Conference on Artificial Intel- Cambridge, England).
Irgence, 366-372.
Clemen, R.T.. 1989, “Combining forecasts: A review and an
annotated bibliography”. International Journal of Forecast-
ing 5. 559-584.
The Economist, 1989, “Management brief: Decisions, deci-
sions”. July 22, 1989, pp. 64-65. Biographies: Bruce ABRAMSON is an Assistant Professor of
Edwards, W.. 1962. “Dynamic decision theory and probabilis- Computer Science at the University of Southern California and
tic information processing”, Human Factors 4, 59-73. a Research Associate at USC’s Social Science Research In-
stitute. He received his B.A., M.S.. and Ph.D. in Computer
Edwards. W.. H. Lindman and L.J. Savage, 1963, “Bay&an
Science, all from Columbia University, and spent two years as
statistical inference for psychological research”, Psychologi-
a researcher at UCLA. His current research combines elements
cal Review 70(3), 193-242. of artificial intelligence, decision analysis, forecasting, diagno-
Einhorn. H.J. and R.M. Hogarth, 1982, “Prediction, diagnosis, sis, probability and statistics. He is particularly interested in
and causal thinking in forecasting”, Journal of Forecasting applications that deal with energy and the environment. Some
1. 23-36. of his recent publications have appeared in ACM Computing
Energy Modeling Forum, 1982, “World oil”, Technical Report Surveys, IEEE Expert, and IEEE Transactions on Pattern
EMF-6, Stanford University. Analysis and Machine Intelligence.
Gately. D.. 1984, “A ten-year retrospective: OPEC and the
world oil market”, Journal of Economic Literature 22, Tony FINIZZA is the Chief Economist for the Atlantic Rich-
llOc-1114. field Company. He is responsible for economic and energy
Hertz. D.B.. 1964, “Risk analysis in capital investing”, Harvard forecasting, business and economic environment inputs to
long-range planning, and energy/economic studies. Before
Business Reuiew, 95-106.
joining ARCO, he taught courses in economics at the Univer-
Heckerman. D.E.. E.J. Horvitz and B.N. Nathwani, 1990,
sity of Illinois, DePaul University, and the University of
“Toward normative expert systems: The pathfinder Chicago. He holds a BA in Mathematics and a MBA from the
project”, Technical Report KSL-90-08, Stanford Univer- University of California at Berkeley, and a Ph.D. in Economics
sity. from the University of Chicago.