Vous êtes sur la page 1sur 8

DISTINGUISHED AUTHOR SERIES

Monte Carlo Simulation:


Its Status and Future
J.A. Murtha, SPE, Consultant

Introduction gate the risks. Thus, turnkey contracts, guarantees, insurance,


Monte Carlo simulation is a statistics-based analysis tool that yields locked-in prices, and hedges are instruments of risk management.
probability-vs.-value relationships for key parameters, including oil A portfolio is an aggregation of investments. Portfolio managers
and gas reserves, capital exposure, and various economic yard- mix their prospects to reduce collective risk and enhance ROJ. Opti-
sticks, such as net present value (NPV) and return on investment mization is often taken as maximizing some measure of reward,
(ROI). These probability relationships help the user answer such such as NPV or profit-to-investment ratio, subject to constraints on
questions as "What is the probability that the NPV of this prospect risk. Strategic planning involves portfolio management but may in-
will exceed the target of $1 ,500,000')" or "How likely is it that the clude more intangible aspects of investments, such as the advantage
reserves added from this year's exploration program will fall short of having a presence in a country.
of our planned production?" Monte Carlo simulation is a part of risk For purposes of this paper, risk analysis is any form of analysis that
analysis and is sometimes performed in conjunction with or as an studies and hence attempts to quantify risks associated with an invest-
alternative to decision [tree] analysis. ment. So that we do not beg the question, we need to define risk. By
Putting aside for the moment a description of Monte Carlo simula- risk, we mean a potential loss, and, more generally, loss or gain (i.e.,
tion, the method has attracted its share of critics over the years. Their a change in assets associated with some chance occurrences). To use
comments include "I did this in FORTRAN in 1964. It just never caught the term analysis, we surely suggest that the risk is quantifiable. The
on."; "Why not just add or subtract 10% to the base case?"; "The an- risks in drilling a well include the direct costs of the rig and of other
swer is whatever you want it to be."; " The answer depends on who is goods and services, the possibilities of unscheduled events and the as-
doing the simulation."; "Garbage in, garbage out."; "You never have sessment of their consequences, the possibility of failure (i.e., a dry
enough data."; "A black box, hocus-pocus, that's all it is."; and "It takes hole, a missed target, or an unsuccessful completion), the range of
too long to run enough cases." To some degree, the critics have been possibilities of success, and the chance of a serious mishap. Risks
silenced by the evolution of virtually universal spreadsheet programs, associated with building a gas-fired electric-power generating plant
much faster computers, and relatively simple software to run simula- include the forecasts of gas price (on the cost side) and electric price
tion and process data. Nonetheless, we will address some of the under- (on the revenue side) as well as capital and operating costs, downtime,
lying concerns, but it is necessary to lay some foundation first. and demand. Risk associated with estimating reserves for an explora-
Our objectives are (I) to define Monte Carlo simulation in a more tion prospect include estimation of the geological chance factors.
general context of risk and decision analysis; (2) to provide some
specific applications, which can be interrelated; (3) to respond to Simple Risk Estimates by Use of Probability
some of the criticisms; (4) to offer some cautions about abuses of the Suppose you have drilled 56 wells in a field-development program,
method and recommend how to avoid the pitfalls; and (5) to predict 14 of which were deemed dry holes. What is the chance that the next
what the future has in store. well will be dry? Perhaps 14/56 is a good answer. It depends on
whether there has been a trend to have fewer or more dry holes over
What Is Risk Analysis7 time. In other words, is the success of one well in some way depen-
Although the word "risk" occurs with great regularity these -days in dent on the success of others? Whatever the case, providing an esti-
the petroleum literature, it has not always been fashionable. In its mate for the probability of a dry hole is a form of risk analysis.
60-page Subject Index, the 1989 printing of the 1,727-page Petro- As a more complicated example, consider the estimate for geo-
leum Engineering Handbook 1 contains just one reference to "risk logical success based on chance factors where the probability of
[factor]" in an article about property evaluation. success is the product of the probabilities of having a structure, a
Among the numerous words and phrases associated with risk anal- seal, a reservoir, a source, and timing. Estimating the factors and ob-
ysis are decision analysis, risk assessment, risk management, portfo- taining the product is a form of risk analysis.
lio management and optimization, and strategic planning. In some As the subsequent examples illustrate, however, the specific risk-
contexts, these words are used only in a qualitative sense, but our fo- analysis technique called Monte Carlo simulation offers a flexible
cus is quantitative. and powerful tool to study complex problems. Monte Carlo analysis
Decision analysis, in its broadest form, includes problem identifi- serves as the focal point for the remainder of this discussion.
cation, specification of objectives and constraints, modeling, uncer-
tainty analysis, sensitivity analysis, and rules that lead to a decision. Probability Distributions and Random Variables
Generally speaking, risk analysis and assessment refer to the quanti- Much of risk analysis consists of estimating something with a range
fication of uncertainty, almost always in the context of possible in- of values rather than with a single value. We report that the wildcat
vestments. In the oil and gas business, although much of the analysis well will require between 56 and 87 days to drill instead of saying
might pertain to reserve size, capital cost, production forecasting, it will take exactly 65 days and that the total cost will be between
and the like, the bottom line universally is monetary value. U.S. $4.3 million and U.S. $7.2 million, not exactly U.S. $5.2 mil-
Risk management connotes a second stage, where the investors lion. Instead ofthe single-point estimate of U.S. $34 million, we say
seek protection from unfavorable situations; i.e., they work to miti- the waterflood prospect's NPV is a normal distribution with a mean
of U.S. $34 million and a standard deviation of U.S. $1.7 million.
Copyright 1997 Society of Petroleum Engineers Thus, as we move toward risk analysis, it becomes necessary to de-
This paper is SPE 37932. Distinguished Author Series articles are general, descriptive rep- velop some familiarity with the rudiments of probability and statis-
resentations that summarize the state of the art in an area of technology by describing recent tics. For our purposes, we need the paired ideas of probability dis-
developments for readers who are not specialists in the topics discussed. Written by individu-
als recognized as experts in the area, these articles provide key references to more definitive tributions and random variables. A random variable is a variable or
work and present specific details only to illustrate the technology. Purpose: to inform the gen parameter that can be described by a probability distribution, which
eral readership of recent advances in various areas of petroleum engineering. A softbound
anthology. SPE Distinguished Author Series: Dec. 19B1Dec. 19B3. is available from SPE's
in tum is an xy plot relating probability to value. There are two com-
Book Order Dept. mon formats for probability distributions: a cumulative distribution

.JIYf April 1997 361


probability vs. cumulative-value relationship. Thus, we may assume
NormalCDF that the area, A, can be described by a lognormal distribution with a
mean of 2,000 acres and a standard deviation of 800 acres, having a

:s~ 0.80
1.00~ practical range of approximately 600 to 5,000 acres. Fig. 2 identifies
.
!'
and shows similar distributions for each of the other input parameters.
11 0.60 -- ~-t-i--~te---t---j A trial consists of randomly selecting one value for each input
o. 40f-----t--- and calculating the output. Thus, we might select A = 3,127 acres,
h = 48 ft, = 18%, S'" = 43%, and Bo = 1.42. This combination of
~ 0 . 2aL-.._--B<=;~----'L..L-. __ -'---_~ values would represent a particular realization of the prospect yield-
(j o. 00 ing 63.5 million bbl of oil. A simulation is a succession of hundreds
a 200 400 600 800 or thousands of repeated trials, during which the output values are
stored in a file in the computer memory. Afterward, the output val-
x ues are diagnosed and usually grouped into a histogram or cu~u~a
PDF Normal Distribution tive distribution function. Fig. 3 shows the output and the sensItIvity
chart for this model.
0.004 T
~0.003t PP~
Selecting Input Distributions. Lognormal distributions are often
used for many of the volumetric model inputs, although net-to-gross
.CI I
ratio and hydrocarbon saturation are seldom skewed right and al-
as 0.002 T )Zl

~ O.OO~ +t. . . . j:O['l<J<;:;?!+.J1_,1ZI_-+-1


'f~, _ _ ==;~_-j
ways sharply truncated. Triangles are also fairly common and are
easy to adapt because they can be symmetric or skewed either left
+--_'
or right. 6 Sometimes, the distributions are truncated to account for
natural limits (porosity cutoffs, well spacing). When all the inputs
o 200 400 600 800 1000
are assumed to be lognormal, with no truncation, and independent
x of one another, the product can be obtained analytically.

Fig. 1-CDF and PDF for normal distribution having mean of 450
Shape of Outputs. In this example, regardless of the distributi?n
and a standard deviation of 117 (P10 =300, Pgo=600). Also types of the inputs, the output is approximately lognormal. That IS,
shown is a Monte Carlo sample. the reserves distribution is always skewed right and "looks" lognor-
mal. In. fact, a product of any kind of distributions, even with
function (CDF), which is a nondecreasing (or non increasing, depend- skewed-left factors, has the approximate shape of a lognormal dis-
tribution. Fig. 3 displays the best-fitting lognormal curve overlaying
ing on preference) function whose y values range from 0 to I. It is sim-
the output histogram for our first example.
plest to think of a random variable as a graph of cumulative probabili-
ty (on the V axis) vs. value. The companion plot, called a probability
density fu'nction (PDF) is simply the derivative of the CDF. Fig. 1 Sensitivity Analysis. One of the byproducts of the simulation is a
sensitivity chart (Fig 3). A measure of significance toward a given
shows both these plots for a normal distribution. Although the CDF
output variable is calculated for each input variable, namely the (rank-
presentation is ultimately more useful, the PDF may be more familiar.
order) correlation coefficient between the two parameters. When the
Once a variable, X, is described by its CDF, we can obtain a Monte
correlation coefficient is close to 0 (as in the case of Bo ), the output
Carlo sample (Fig. I) by first generating a random value, Y, uniform-
(reserves) can achieve a value near the upper limit of its range with
ly distributed between 0 and I, then by taking the inverse function
a small Bo and vice versa. The fact that it is negative here is trivial;
of that value from the CDF curve to get X. Simulation software 2.3
there is essentially no correlation between Bo and N. These coeffi-
does all this behind the scenes; the user's job is to specify appropri-
cients can range from - 1.0 to 1.0. Area is identified as the most sig-
ate probability distributions.
nificant parameter, and the others are ranked accordingly. As models
become more complex, sensitivity analysis identifies the "driving
Percentiles
variables" that merit additional scrutiny and, by contrast, helps reduce
Common use of a CDF is to specify certain values of a variable. X. The
effort wasted on worrying about the wrong things. Unlike the tradi-
expression, PIO, meaning the 10th percentile, is the value of X corre-
tional tornado chart or spider diagrams obtained by tracking the
sponding to 0.10 on the cumulative-probability axis. Experienced users
changes in an output caused by allowing exactly one model input to
often describe distributions by specifying two or three percentiles. Of
vary while holding the others fixed, this sensitivity analysis from
particular interest is Pso, which is called the median and is one of the
Monte Carlo simulation is far more versatile because it permits func-
common "measures of central tendency," like the mode (most popular
tional or correlation-type relationships among the inputs.
range; most likely category) and mean (arithmetic average value).
Correlations Among Inputs (Dependency). Unless otherwise spe-
Risk Analysis in the Form of Monte Carlo Simulation
cified, the samples of the various input distributions are taken inde-
Monte Carlo simulation is an alternative to both single-point (determin- pendent of one another. Thus, on a particular trial, a very large value
istic) estimation and the scenario approach that presents worst-, most- of Sll' may be paired with a very large value of or a small value of
likely-, and best-case scenarios. The term Monte Carlo dates back to A could bc chosen along with a large net pay. If there is a geological
the Manhattan Project in the 1940's, where it was used as a code name. argument (or empirical data) to justify a correlation between ~ pair of
The inference is to the gambling mecca, where chance rules. For an ear- parameters, then the simulation software can honor that relatI<;mshlp.
ly historical review, see Ref. 4. The following paragraph, drawn largely Correlation among inputs, in this particular example (not In gen-
from Ref. 5, offers a brief description of the Monte Carlo process. eral) causes the resulting output to become more dispersed. Thus,
A Monte Carlo simulation begins with a model (i.e., one or more when area and pay (often) or porosity and water saturation (nearly
equations, together with assumptions and logic relating the parame- always) are correlated, the prospect appears to be riskier and the
ters in the equations). For purposes of illustration, we select one form mean value of the estimate increases somewhat when the correla-
of a volumetric model for oil in place, N, in terms of area, A; net pay, tions are incorporated in the model. Sometimes, the difference is
h; porosity, ; water saturation, SI\'; and formation volume factor, Bo. slight, but the modeler should be prepared to describe at what level
N = 7, 758Ah(I-S.,)/Bo . ..... , .................. (I) the correlations make a difference. 5 Imposing correlations among
inputs can also reorder the list of sensitive variables.
Think of A, h, ' SI\" and Bo as input parameters and N as the output.
Once we specify values for each input, we can calculate an output val- Model Shortcomings. There is a widespread practice of extracting
ue. Each parameter is viewed as a random variable; it satisfies some one or a few numbers from the reserve distribution to "run econom-

362 April 1997 ,JPT


---------
Area 1 Forecast: Area
------------------------------, I
[1,000 Trials Cumulative Chart 11 Outliers
1.000 i------- -----=:::::;;:::==Jrr 1000
~750 ~
... ! 150
:
:tI
::=500--- _ 500 -="
I: a= i
584.60 1.913.08
-------.-,.
3,241.55 4.57002 5.898.49
:~
,
.250_

_______ -=__
~.----""-~---t-'-:.
.000 --.----

___'.500
'[~:~

______2_.500
____3._500
_ _ _,_.500_
250

~
.;,e I

Lognonnal, mean 2000, STD 800 acres


pay
.-----------------------,
1
I
I

36.00 -40.50 45.00 411."


J
5400

Nonnal, Mean 45, STD 3 feet

Triangular, 0.20, 0.30, 0.45 fraction


Sw

0.20 0.26 0.33 0.39 0.45

Fig. 2-Five input distributions, with area shown in both CDF and PDF formats; STD = standard deviation.

ics," treating well productivity and capital costs as deterministic and Monte Carlo Simulation-Typical Problems
thereby reducing the effectiveness of the uncertainty analysis. The In addition to the widely used volumetric-reserves model, the fol-
field-development model that follows picks up where the reserves ex- lowing three problems serve to illustrate the breadth of Monte Carlo
ample leaves off. As engineers and geoscientists become more accus- simulation. Taken as a whole, they also suggest a process of integra-
tomed to integrating risk-analysis components, the exploration-pros- tion or consolidation.
pect model will include scoping estimates for production forecasts
and both capital and operating expenses. Without this integration, Problem 1. Drilling Authorization for Expenditure (AFE): Esti-
identifying the relative importance of the various risks is difficult. mate Total Costs and Times. Drilling APE estimators are prime can-
How does field-size uncertainty compare with the uncertainty of the didates for Monte Carlo simulation, being conceptually simple, ubiqui-
capital costs or the productivity of the wells? Where should we invest tous, and essential to the overall prospect evaluation process.7-9 Fig. 4
additional resources to acquire information? What are the key issues shows a simple format for the AFE model, with categories of activity
when negotiating with partners or forming alliances? These questions times and costs of goods and services. One task for the user is to esti-
are not easily answered, yet they should be acknowledged.
mate durations for various activities, such as drilling the hole sections,
completing, and testing. Another task is to estimate line-item costs. In
Why Do Monte Carlo Simulation?
this model, the drilling expert provides two values for each estimate:
We do risk analysis because there is uncertainty in our estimates of
capital, reserves, and such economic yardsticks as NPY. Quantify- a low and a high estimate. A preliminary calculation in the worksheet
ing that uncertainty with ranges of possible values and associated solves some simple equations 10 to obtain the mean and standard devi-
probabilities (i.e., with probability distributions) helps everyone un- ation for a lognormal distribution for that cost or time category. These
derstand the risks involved. There is always an underlying model, high and low estimates are treated as PIO and P90: the actual value has
such as a volumetric reserves estimate, a production forecast, a cost a 10% chance of being less than the low value and a 10% chance of
estimate, or a production-sharing-economics analysis. As we inves- exceeding the high value. One could treat the low estimate as a P5 and
tigate the model parameters and assign probability distributions and the high estimate as a P95, or variations thereof. Because the input pa-
correlations, we are forced to examine the logic of the model. rameters represent times and costs, they are generally regarded as
The language of risk analysis is precise; it aids communication, skewed right; there are more extreme values to the right of the mode
reveals assumptions, and reduces mushy phrases and buzz words. than to the left, causing a tail of large values having low probability of
This language requires study and most engineers have little expo- occurrence. Alternative distributions include beta, gamma. and triangu-
sure to probability and statistics in undergraduate programs. lar (which would require three user estimates).

364 April 1997 JIYf


SensItivity CharI ~t=
Generic AFE Model
:_==:::--:.:::::..::. ._. ---- .- ~J File: AFEGENXts

-
~_: Murth:;;l994
TBIlI8I FonIcasI: Reserves
..... .84
IUgDayRate. $M 110
Cost Estimates
Sw .35


"....,~ .30
pay .15
eo -.05

-0.5 05
Measured bv Rank Coffelilllon
Overlay Chart
Reserves
.105--'- Rig=TC---
e1 im
+-~D~e-sc-riP-tti-on---fO'Ti,..m=ce~~SoC;tim
...a-::te~~=-c-'~E;~ -- ----~.~---.-
.078 _ _ _ _ _ _ _ lognormal DIstribution
Mob/Demob 0.5 2.1: 1.20-13'2 --132 ----
Mean .23
~1R-30;';-'- 2.5' 2~i4 --247----247----
SIdDev"11 ~~ __ ~~i~ ____ 4 7 5.421 596 596
Stage3 13-3/8 in 9 12 10.46 1150 1150
12 25 18.05, 1985
- - - - - - - - - -: Reserves 7 13 9.82 1080
Stage6 0.00' o
.000 .:_

I. ,. 55
"
Fig. 3-0utput distribution (histogram) with fitled lognormal
curve. Fig. 4-Drilling AFE estimator template.

Acknowledging Historical Data. When data are available from the capital cost of drilling and completing the well. Note that this
analogous projects, one can fit particular distributions to specific in- distribution should come from the drilling engineers who model the
put parameters. There is a tendency for engineers to want to lean AFE simulators. Further assume that operating costs can be de-
heavily on data, sometimes at the risk of including inappropriate data. scribed in terms of probability distributions. We might want to break
Certainly, in time, more and better data may be available. One must cost into fixed and variable components and to treat variable costs
often settle for carefully designed sensitivity analysis to quantify the as being proportional to oil production or water production. Alterna-
drivers and the impact of changing distribution types and ranges. tively, we could correlate operating expense and oil production.
Generate the forecast of production and discounted cash flow for
Warning. Do not report partial results. Too often, the results of a sim-
ulation are reduced to a mean value or a P50. Sometimes two or three
values, such as PIO, mode, and P90, are passed along to the next level,
where they are used to "run economics." Such practices are not in the
spirit of Monte Carlo simulation. As described later, the economics
itself must be a simulation that samples from the full distributions for
the drilling cost as well as from distributions for prices, expenses, and
production forecasts. One cannot capture PIO-case economics run by
splicing together PIO cases for the numerous ingredients.
The proper role for the drilling AFE is a preliminary step,to more
comprehensive models. The outputs (total time and costs) can be re-
ported as a distribution or a series of distributions over time or for
each of several wells. In cases where there are multiple wells ordrill-
ing costs along with facilities costs, care must be taken to transmit
the correlations between the relevant distributions.

Other Cost-Estimation Models. Aside from the utility of the AFE


model in its own right. it serves as a paradigm for other cost-estima-
tion models. Many cost engineers use Monte Carlo techniques for
pipelines, platforms, and other facilities. Even the notions of contin-
gency can be described in useful terms. For purposes of including
them in the Monte Carlo economics model, simple spreadsheet line-
item models, similar to the drilling AFE, are adequate. When sched-
ule details are paramount, a project-scheduling model is necessary.
In recent years, Monte Carlo versions of the most popular schedul-
ing programs have become available.

Problem 2. Field-Development Program: Estimate Production


and Revenue Streams. Start with a deterministic forecast of oi I pro-
duction, possibly generated by an exponential-decline model,
q = qjexp(- at), ............................... (2)

where qj = initial rate, stock-tank barrels per year; a = fractional de-


cline per year; and t = time in years. Think of both qj and a as proba-
bility distributions. We do not know exactly how much oil will be
produced in the first year or how sharply the production will decline
from one year to the next. Likewise, assume some distribution for

JlYC April 1997 365


FS:::'r.:::'e=gi'c-'-P:::I.n"-ETx=pll,o:..::r.=tiornc:.Pro.:.:o=-;gr'amr-_--EF~i1e~:",CB",M U",LTP2.XLS - - - - f - - - - - - j - - - - - j
7
(c) JimMurij,.,j99
r_ ~-
Reserves/llisCo~- ~lbiSc:;;-=-
Trend Chart
Inp:uts
4.5,-_ _ _ _ _ _ _ _ _ _ _ _ _ _ _--,.. PIS) Wells Dry Hole Cos,
fraction number $MM MMBOE $MM

~~;;:--------_t 0 900/0 2.1 5.5 88 320


Rate 3.9
4.4
7.7
9
90
200
430
600
-- -- - -

1.1t----~~~o;::__---__+ III 50% Simulation Results

Success Capital Reserves NPV


---------t===-+.;;$M~M:=-=--1E~~~~+---+-----+--+-----l
MMBOE $MM
Pdcili-Z-run;---
3 219 194 334
~Alnenc.~-o --329 ----~;;i+_-~ct-----t---+---+---j
Time Nor,h Se. ~-1 --T7O -166
108 342
301
~~ __ ~-656 _.E! 396
1------
- -- - - - -
To'als - - 4 18:74 742 1304
Fig. 5-Production forecast with uncertain ramp-up, peak rate,
and two decline segments. Fig. 6-Spreadsheet template for exploration portfolio.

some time horizon (e.g., 20 years). Calculate NPV at 10% interest. mon price and cost forecasts. The resulting simulation results would
So far, we have a single-well exponential-decline forecast for a yield the desired correlations.
single production stream.
This model generalizes to any number of wells. The underlying Scenario Errors
deterministic production forecast need not be exponential decline. Deterministic reserves estimates are often described in terms of
Fig. 5 illustrates a more complicated production profile with a delay low-, medium-, and high-side possibilities. Some people think in
and ramp-up followed by two declines. Wells can be linked by cor- terms of extremes: worst and best cases together with some base
relations or other relationships. More than one zone in a well can be case (mean, mode, or median). Others report PlO, Pso, and P90 val-
modeled, and associated gas production can be accounted for in the ues. Sometimes, these cases are linked to such categories as proved,
model. Workovers, waterfloods (i.e., capital investments, with un- proved' plus probable, and proved plus probable plus possible.
certainty and future changes in production streams) are possible. While there is nothing wrong with any of these notions, the logic of
Anything that can be handled deterministically can be included in obtaining the cases is often flawed.
the stochastic model.
The output distributions of this model are the requisite input dis- Do Not Mulitply or Add PlO Values. When volumetric products
tributions for portfolio-optimization models. Namely, these are fore- are used to obtain reserve estimates, there is a temptation to build the
casts (1) of aggregate field production by type, which sum to "re- low-side reserve estimate by taking the product of low estimates for
serves"; (2) of cash flows, one of whose sums is NPV, and ROI; and the various factors. This is a dangerous business at best. The product
(3) of capital and operating expense, providing a measure of "expo- of PlO estimates for area, pay, and recovery factor, for example, is
sure." The usefulness of this model is obvious. One can estimate the approximately PI. For special cases (all distributions lognormal, no
probability of achieving some hurdle rate for ROJ or of failing to meet correlations), one can find an exact answer. But if you use a different
a target NPY. The sensitivity analysis highlights the input parameters distribution type, include any correlations between inputs (larger
that drive the model and alerts investors to opportunities to mitigate
area tends to be associated with thicker pay), or change the number
the risks. ROJ is a parameter that needs special attention when doing
of factors (breaking out recovery factor into porosity, saturation,
a simulation because it relies on an iterative process to converge and
formation volume factor, and efficiency), there are no simple rules
may occasionally fail. The resulting ROJ output distribution then has
of thumb to predict just how extreme the product of PlO values is.
some missing data. There are several ways to avoid this pitfall.

Problem 3. Strategic Plan: Estimate Reserve Increases, NPV, Do Not Multiply Pso Values or Modes. Less obvious is the fact that
and Capital Exposure. The strategic-plan model can be rather sim- the Pso value or the modes for the inputs do not yield Pso or mode,
ple. The objective is to aggregate various capital investments (let us respectively, for the output, except in very special cases. The mean
call them ventures) making up a portfolio. Jn the upstream side of values of inputs will yield the mean of the product distribution, pro-
the petroleum industry, these ventures might represent countries. vided there is no correlation among inputs. In other words, even the
Each investment is represented by distributions for capital, reserves, "base-case" reserve estimate generally should not be obtained from
and NPY. Restricting the portfolio further to an exploration portfolio a product of base-case inputs except in very special cases.
might add the number of wells planned and success rates, along with Similar arguments apply to cost models, such as a drilling AFE
the associated estimates for reserves, cost (both exploration and de- model, where the principal output is a sum of inputs. One cannot
velopment), and NPY. Figs. 6 and 7 show the spreadsheet template simply add the low and high estimates for the line items and hope
and typical inputs and outputs for such a model. to obtain realistic low and high estimates for the total. Typical sums
On each trial of the simulation, samples are taken from number of of PIO values might be PI or P2, and could be much less likely (see
discoveries for each venture, along with corresponding reserves, Fig. 8). Again the exact value of the sum of PIO values depends on
NPV's, and capital exposures (both exploration and development). distribution types, relative sizes of inputs, and correlations. It is im-
Outputs are the aggregations of these parameters. Aggregation mod- prudent to try to generalize. Fig. 8 illustrates a postsimulation graph
els rely on good input distributions and any correlations between of the Ps and P9S curves for depth vs. time. The obviously incorrect
them. These input distributions are themselves outputs from an explo- curves representing the sum of the best and worst values for the in-
ration-economics model. Correlations arise from technology changes puts are included. This is the analogy of obtaining low and high esti-
(high productivity in several wells might depend on state-of-the-art mates of reserves by taking products of PIO and P90 values.
drilling or completion), economic factors (changes in prices of prod-
ucts and in costs of goods and services affect all projects), and estima- Do Not Add Pso Values or Modes. Worse yet, one cannot simply
tor biases (capital costs estimates on various projects might be made add modes or Pso values from individual cost models to obtain the
by the same team, which tends to underestimate certain line items). mode or Pso, respectively, of an aggregation. Only means add to
Estimating the coefficients of these correlations may be difficult. One means. Yet, it is a common practice to estimate the base cost of a
can (and indeed should) run the various economic models using com- drilling program by adding the base estimates for individual wells.

366 April 1997. JPT


Forecast: Wells Forecast: Capital
.- 1_ ... 1_ '00

~ /'
....i
.1"
.1" ~1I
I
... .."""
if ~
~ j .... J
:iii .... .a I:

.a
r'
I:

"
.a
~ / ~"
~
.210 121
l /
.- ,
.2M 121

~
.100

....

Forecast: Reserves
11
I'
.... ,... Forecast: NPV -
AU ".1 .....

'III
/' n.
""
~ ,n
/

/
J ,..
i" f ... '50 t
;'
Il. ~ i ... / u. ~
II

om / ... /
~
I", ''''' .... ~
"....
...""" '.110 '3.6111
"'M
15

Fig. 7-0utputs for exploration-portfolio model.

Similarly, total capital cost is often estimated by adding the base been pockets of resistance. These efforts sometimes date back to the
costs for the various line items. 1970's, some even further. Since Hertz'sll widely read article, there
A simple exercise shows how far off the total cost can be. Take 10 have been the sporadic articles 12 -28 and books 29 -34 on Monte Carlo
identical triangular distributions, each having 100. 200, and 350 for simulation and related topics in the petroleum literature. The attention
low, most-likely, and high values, respectively. While the mode of
each is 200, the mean is 216.7 The sum of these distributions is
approximately normal with a mean of 2, 167 and a standard deviation
of approximately 18. The sum of modes, 2,000, is well outside the 3-
standard-deviation range. In other words, 10 independent line items,
each ranging in cost from 100 to 350 with most likely of 200, would
have an aggregate cost in the range 2,113 to 2,221. Do not even con-
sider adding the high or low estimates. The modes themselves sum
to values outside the range of practical consideration.

What Monte Carlo Simulation Does Not Do


In spite of its power and applicability, Monte Carlo simulation
does not do the following.
I. It does not make decisions; it prepares for decision making.
2. It does not analyze data; there is companion software for that
purpose.
3. It does not optimize functions; the output distributions serve as
ingredients for optimization.
4. It does not provide ready-made models; everyone builds their
own.

Who Does Risk Analysis in the Oil and Gas Industry?


Virtually all segments of the industry-operating companies, large
and small; service companies; consulting firms; and financial insti-
tutions-are engaged in some form of risk analysis. Thousands of
geoscientists and engineers have attended classes in probabilistic
reserve estimation, general principles, or hands-on Monte Carlo
simulation: and the pace of that training is accelerating. Several ma-
jors have organized some sort of process to do risk or decision analy-
sis on a systematic basis. They have created task forces, organized
procedures, held in-house training. molded internal specialists. and
designated champions. Seldom, however, has this been attempted
at a corporate level. Rather, it may be done in the exploration group
or within the planning department or by the drilling engineers.
At this time, it appears that no oil and gas operating company has
a fully implemented, unified program to do probabilistic estimates of
their key operating parameters: reserves. capital exposure. and NPY.
A few have made efforts in this direction; there always seems to have

.Jly (' April 1997 369


given to risk analysis in recent years should begin to generate more
research and broader literature.

Responses to Criticisms of Monte Carlo Simulation Drilling Performance:


Many of the criticisms of Monte Carlo simulation fall into the fol-
lowing categories.
1. It didn't work 20 years ago.
2. There are no adequate data.
3. Outputs depend on people running the models.
In response, one can cite the development of spreadsheets and fast
computers. Instant graphics and statistical analysis, simple pro-
gramming, row and column structure, hundreds of canned func-
tions, and an estimated 30 million official users make Excel and Lo-
tus 1-2-3 natural vehicles to do Monte Carlo simulation.
Data can be a problem; but either you have them and can easily ob-
tain statistical curve fits or you don't (often the case) and you learn -18000 +----1----1"----+----
how to do sensitivity testing to find out the key parameters and how
they affect results when the distribution type and ranges are altered. -22000
I
1 - - __

Then, you make a concerted effort to acquire data in the future. The
point is this. If you can make a deterministic estimate for some param- Time, day
eter, then you can make an estimate for its possible range (e.g., from
PIO to P90). Many people feel more comfortable presenting the range.
Item 3 is addressed in the guidelines in the next section. Suffice i ~;~-;-Pl0--~-P5-- -Sampled I
it to say, there are many abuses of Monte Carlo simulation (and of
decision trees). One example is multiple models within a company 1--><- P95 ---"_s_um_p"_90_ _ _ _----11
to estimate the same thing. There should be exactly one model for
charge assessment, one for reserves estimating, one for drilling AFE
model, and one standard facilities cost model. Like any other soft-
ware tool, they should be designed carefully with input from the key
users, maintained over the long term, and reviewed periodically.
Fig. 8-Time-vs.-depth forecast with actual Ps and P9S vs. sums
of P10 and Pgo inputs.
Some Gui~elines for Monte Carlo Simulation
I. Establish comprehensive objectives: scope, deadlines, process.
3. Some form of geostatistics interface to Monte Carlo simulation.
2. Hold awareness seminars on the language concepts of risk
4. Formalized agreements for sharing data and model description.
analysis.
5. Renewed attempts at systematic integration of risk and deci-
3. Design and conduct training in Monte Carlo software.
sion analysis into companies.
4. Decide which models to build or buy.
5. Designate teams or indi viduals to design and acquire the models.
Future Developments.
6. Establish reporting procedures: forms, timing, and people.
I. Commercial Monte Carlo templates for economics, drilling!
7. Keep reports of results to a few pages and 30 minutes. However,
facilities costs, and production forecasts.
insist that they include (1) sensitivity analysis with 10 or fewer driving
2. Seamless links between components, with empirical data.
input parameters, (2) the basis of selection of all key inputs, (3) identi-
cal-format output distributions, and (4) signatures of reviewers. 3" Several corporations embracing the method.
8. Avoid duplication or modification of models; use periodic re- 4. Several portfolio and planning models; at least two commer-
view to suggest changes. cial models already exist.
9. For each model, identify two people who will be responsible 5. User groups for Monte Carlo software.
for explanations, assistance in running, review, and collection of re- 6. Look-backs; companies will formalize this process.
usable examples and who will, in general, assume responsibility for 7. Catalogs of distribution types.
the integrity of the model and its use. 8. Steady growth of technical papers.
10. Set firm guidelines for documenting all models and applica-
tions religiously and thoroughly. The Perfect World. The following would be present in an ideal en-
11. Establish look-back procedures that focus on improvement of vironment for engineers and scientists
estimations and of the model logic, not on reward or punishment. I. The language of basic statistics would be as common as lan-
12. Always ask whether there is another model doing essentially guage of engineering economics.
the same thing and offer an incentive to one of the model builders 2. Appropriate databases would be generated, properly main-
to join your company. tained, and used to guide parameter modeling.
13. Periodically have a disinterested third-party review the ap- 3. Everyone would know the significance of their analyses and
plications. how their results plugged into the next level of model.
4. All estimates would be ranges; single-number requests would
Future of Monte Carlo Simulation be refused.
Is there a rebirth of simulation, and will it persevere? Here are some 5. Budgets would be built on distributions. Aggregation, properly
predictions. done, results in estimates with relatively less dispersion than the in-
dividual components. We too often fool ourselves with single num-
Near-Term Developments. bers and then either force spending or create reasons for missing the
1. Field testing of two commercial economics packages inter- target. There would be no penalty for coming in "over budget." Per-
faced to Monte Carlo spreadsheets. formance measures would be probabilistic.
2. Testing of at least two stand-alone comprehensive Monte Carlo Can this happen? Of course! We must assume that everyone can
economics-evaluation models. be educated. .JPI'

370 April 1997 n ....


References 29. Newendorp, P.: Decision Analysis for Petroleum Exploration, Penn Well
I. Bradley, H.B.: Petroleum Engineering Handbook, SPE, Richardson, Publishing Co., Tulsa, Oklahoma (1975).
Texas (1987). 30. McCray, A.W.: Petroleum Evaluations and Economic Decisions, Pren-
2. @RISK-Adl'anced Risk Analysisfor Spreadsheets, User:~ GlIide, Pali- tice-HaIl Inc., Englewood Cliffs, New lersey (1975).
sade Corp., Newfield, New York (1996). 31. Megill, R.E.: An Introduction to Risk Analysis, Petroleum Publishing
3. Crystal Ball Users Manual, Decisioneering, Aurora, Colorado (1996). Co., Tulsa, Oklahoma (1977).
4. Halton, J.H.: "A Retrospective and Prospective Survey of the Monte Car- 32. Clemen, R.T: Making Hard Decisions-An Introduction to Decision
lo Method," SIAM Reviews (1970) 12, I. Analysis, PWS-Kent Publishing, Boston, Massachusetts (1991).
5. Murtha, lA.: "Incorporating Historical Data in Monte Carlo Simula- 33. Murtha, J .A.: Decisions Involving Uncertainty-An @RISKTutorialfor
tion," SPECA (April 1994) I I. the Petroleum Industry, Palisade Corp., Newfield, New York (1993).
6. Dhir, R., Dern, R.R. and Mavor, M.1.: "Economic and Reserve Evalua- 34. Harbaugh, J., Davis, J., and Wendebourg, 1.: Computing Riskfor Oil
tion of Coal bed Methane Reservoirs," lPT(December 1991) 1424. Prospects, Pergamon Press Inc. Elmsford, New York (1995).
7. Peterson, S.K., Murtha, J.A., and Schneider, F.F.: "Risk Analysis and
Monte Carlo Simulation Applied to the Generation of Drilling AFE Esti- SI Metric Conversion Factors
mates," paper SPE 26339 presented at the 1993 SPE Annual Technical acre x 4.046 873 E-OI =ha
Conference and Exhibition, Houston, 3-6 October. bbl X 1.589 873 E - 01 = m3
8. Peterson, S.K., Murtha, J.A., and Roberts, R.W.: "Drilling Performance ftx3.048* E-Ol=m
Predictions: Case Studies IIIustratiog the Use of Risk Analysis," paper
in. X 2.54* E+OO=cm
SPE 29364 presented at the 1995 SPE/IADC Drilling Conference, Am-
sterdam, 28 February-2 March. "Conversion factor is exact.
9. Kitchel, B. et al.: "Probabilistic Drilling Cost Estimating," paper SPE
35990 presented at the 1996 SPE Petroleum Computer Conference, Dal-
las, 3-5 June. James A. Murtha is an independent consultant
10. Murtha, J.A. and Janusz, G.1.: "Spreadsheets Generate and Validate Un- specializing in risk analysis and decision-mak-
certainty Distributions," Oil & Gas 1. (13 March 1995) 87. ing. He holds a BA degree from Marietta C. and
I I. Hertz, D.B., "Risk Analysis in Capital Investments," Han'ard Business MS and PhD degrees from the U. of Wisconsin,
Review (January-February 1964) 95. all in mathematics, and an MS degree in petro-
12. Walstrom, J.E., Mueller, TD., and McFarlane, R.C.: "Evaluating Uncer- leum and natural gas engineering from Pennsyl-
tainty in Engineering Calculations," lPT(December 1967) 1595; Trans., vania State U. Murtha is an SPE Short Course
AIME,240. instructor and a member of the Forum Series in
13. Smith, M.B.: "Probability Estimates for Petroleum Drilling Decisions," Asia Pacific Steering Committee. He was a 1995-96 Distin-
lPT (June 1974) 687: Trans., AIME, 257. guished Lecturer and a 1994 Speakers Bureau speaker. He
14. Eraillating & Managing Risk-A Collection of Readings, R.E. Megill chaired the 1994 SPE Petroleum Computer Conference Pro-
(ed.), SciData Publishing, Tulsa, Oklahoma (1985). gram Committee and has served on the Electronic Publishing,
IS. Harrell, lA.: 'The Analysis of Bivariate Association," Use and Abllse {If Computer Applications, and Editorial Review committees.
Statistical Methods ill the Earth Sciences, W.B. Size (cd.), Oxford U. Press,
Oxford, U.K. (1987).
16. Newendorp, P. and Quick, A.N.: "The Need for More Reliable Decision
Criteria for Drilling Prospect Analyses," paper SPE 16318 presented at
the 1987 SPE/SPEE Hydrocarbon Economics and Evaluation Sympo-
sium, Dallas, 2-3 March.
17. Murtha, 1.A.: "Intill Drilling in the Clinton: Monte Carlo Techniques Ap-
plied to the Material Balance Equation," paper SPE 17068 presented at
the 1987 SPE Eastern Regional Meeting, Pittsburgh, Pennsylvania,
21-23 October.
18. Behrenbruch, P. et al.: "Uncertainty and Risk in Petroleum Exploration
and Development: The Expectation Curve Method," paper SPE 19475
presented at the 1989 SPE Asia-Pacific Conference, Sydney, 13-15
September.
19. Damsleth, E. and Hage, A.: "Maximum Information at Minimum Cost:
A North Sea Field Development Study Using Experimental Design,"
lPT (December 1992) 1350.
20. Shivers, R.M. and Domangue, R.1.: "Operational Decision Making for
Stuck-Pipe Incidents in the Gulf of Mexico: A Risk Economics Ap-
proach," SPEDC (June 1993) 125.
21. Davies, G.G., Whiteside, M.W., and Young, M.S.: "An Integrated Ap-
proach to Prospect Evaluation," paper 231 57 presented at the 1991 SPE
Offshore Europe Conference. Aberdeen, 3-6 September.
22. Caldwell, R.H. and Heather, D.I.: "How To Evaluate Hard-To-Evaluate
Reserves," lPT(August 1991) 998.
23. Cronquist, c.: "Reserves and Probabilities-Synergism or Anachro-
nism?," lPT(October 1991) 1258.
24. Hightower, M.L. and David, A.: "Portfolio Modeling: A Technique for
Sophisticated Oil and Gas Investors," paper SPE 22016 presented at the
1991 SPE Hydrocarbon Economics and Evaluation Symposium. Dallas,
11-12 April.
25. Holtz, M.H.: "Estimating Oil Reserve Variability by Combining Geolog-
ic and Engineering Parameters," paper SPE 25827 presented at the 1993
SPE Hydrocarbon Economics and Evaluation Symposium, Dallas,
29-30 March.
26. Edwards, R.A. and Hewett, TA.:"Quantification of Production Uncer-
tainty and its Impact on the Management of Oil and Gas Price Risk," pa-
per SPE 28330 presented at the 1993 SPE Annual Technical Conference
and Exhibition. Houston, 3-6 October.
27. Gutleber, D., Heiberger, E., and Morris, T.: "Simulation Analysis for in-
tegrated Evaluation of Technical and Commercial Risk," lPT (Decem-
ber 1995) 1062.
28. Murtha, J.A.: "Estimating Reserves and Success for a Prospect With
Geologically Dependent Layers." SPERE (February 1996) 37 .

.JIYf April 1997 373