Download as pdf or txt
Download as pdf or txt
You are on page 1of 32

The Creaming Method: A Bayesian Procedure to Forecast Future Oil and Gas Discoveries in

Mature Exploration Provinces


Author(s): J. Meisner and F. Demirmen
Source: Journal of the Royal Statistical Society. Series A (General), Vol. 144, No. 1 (1981), pp. 1-
31
Published by: Wiley for the Royal Statistical Society
Stable URL: http://www.jstor.org/stable/2982158
Accessed: 13-05-2015 15:49 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

Wiley and Royal Statistical Society are collaborating with JSTOR to digitize, preserve and extend access to Journal of the
Royal Statistical Society. Series A (General).

http://www.jstor.org

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
J. R. Statist.Soc. A (1981),
144, Part 1, pp. 1-31

The Creaming Method: A Bayesian Procedureto Forecast Future Oil and Gas
Discoveriesin Mature ExplorationProvinces

By J. MEISNER and F. DEMIRMEN


Shell International
PetroleumCo.
[Read beforetheROYAL STATISTICAL SOCIETY on Wednesday,July2nd,1980,thePresidentProfessorD. R. Cox inthe
Chair]

SUMMARY
The creamingmethodis a statisticalprocedureaimed at forecasting futureoil and gas
discoveriesin petroleumprovinces,based on trendsin explorationresultsobservedin
thoseprovinces.The modelrecognizesthefactthat,generally, thediscoveryrate(success
rate)and thevolumes(fieldsizes)tendto declinewithadvancingexploration.Thisis called
thecreamingphenomenon.Trendsin successrateand fieldsizes are analysedseparately
and departuresfromthe model are investigated.
The forecastis based on these analyses and uses as an indicatorthe numberof
explorationwellsexpectedor plannedto be drilled.The resultsare givenin theformof
predictiveprobabilitydistributions.
These includeestimatesofthenumberand thetotal
volumeof futurediscoveries.
The methodis applicableto areas wherediscoveriesare on a generallydecliningtrend,
or, forshort-termestimates,on a generallyconstanttrend.
Keywords:DISCOVERY; EXPECTATION CURVE; EXPLORATION; EXPLORATION MATURITY;
EXPLORATION TREND; FORECASTING; GAS; LOGISTIC MODEL; LOGNORMAL
DISTRIBUTION; OIL; PLANNING; PREDICTIVE DISTRIBUTION; PROVINCE; RESERVES;
SUCCESS RATE; STATISTICAL ANALYSIS

1. INTRODUCTION
1.1. General
FORECASTSof futurediscoveries of crude oil and natural gas in petroleum provinces are of great
importance forplanning and policy-makingin the petroleum and related industry.In thispaper
we discuss a relativelysimple statistical procedure to forecastfuturediscoveries in exploration
provinces. By "futurediscoveries" we mean the prospective quantity of crude oil or natural gas
yet to be found by a given number of exploration wells which are planned to be drilled in a
particular petroleum province. The term "petroleum province" is used in a broad sense,
referringto a regional geographic area known to have commerical hydrocarbon accumulations,
and need not referto a geologic province with distinctgeologic (e.g. tectonic) properties or oil-
accumulation conditions. For practical purposes, national frontierscan be considered to be
province boundaries.
Our method is intended basically forpetroleum provinces that are in a mature exploration
stage. For making short-termforecasts,however, the method is also applicable to exploration
provinces that are in a transitional stage from immature to mature. A mature exploration
province is one in which, after drilling a relatively large number of exploration wells, the
discoveries are on a general declining trend.If the discoveries are on a general risingtrend,the
province is immature, and if the discoveries show a general constant trend, the area is in
transition fromimmature to mature.
Furthermore,although our method can be used to forecastfuturediscoveries of both crude
oil and natural gas, forecastsforthe two should preferablybe kept separate. If it is necessary to
include both oil and gas in a forecast,the volumetric units of the two should be compatible, as
for example by expressing gas volumes as BTU-equivalent oil volumes.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
2 MEISNERAND DEMIRMEN - The CreamingMethod [Part 1,
1.2. Past ExplorationTrends
Past explorationtrendsin a given petroleumprovince formthe cornerstoneof our
forecasting procedure.We considerpast explorationtrendsin termsofdiscoveriesmade as a
function ofthenumberofexplorationwellsdrilled.Plottingthecumulativediscoveriesagainst
thetotalnumberofexplorationwellsina historicalsequenceprovidesa good visualdescription
ofsuch explorationtrends(Fig. 1). Some kindofgrowthcurveis thenobtained.For a mature
explorationprovince,the growthrate will be on a decliningcourse and fora provincein a
transitionalstate,itwillbe constant.Forecastingbyour methodis tantamountto projectionof
thepast explorationtrendintothefutureovera specifiednumberofexplorationwellsplanned
to be drilled.The numberof futureexplorationwellsservesas an indicatorforthe forecast.

0 >
0-

PREDICTIVE
_ DISTRIBUTION

PRES. CUM. DISCOVERY G -.-

/ I ~~FUTURE I
I EXPL. TREND I
' (FORECAST)
0

PAST EXPL. TREND l

Ai 8
EXPL. WELLS DRILLED ADD. EXPL. WELLS
TILL PRESENT IN FUTURE

ADVANCING EXPLORATION

FIG. 1. Forecast of futureoil discoveries.

Explorationtrendsofthetypeconsideredabove are thecombinedresultoftwoindependent


factors:(i) success rate, that is the proportionof explorationwells that resultedin field
discoveries, recoverablevolumeofoil or gas attributable
and (ii) fieldsize,thatis theultimately
to a field.In our methodwe considerboth factorsseparately,thoughin thefinalanalysisthe
effectsofthetwoare combinedin makingpredictions. Experiencehas shownthatthefieldsize
playsa fargreaterrolein defining theshape ofthediscoverytrendthandoes thesuccessrate(see
also below).

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNER AND DEMIRMEN- The Creaming Method 3
1.3. CreamingPhenomenon
The statisticalmodel we employ for predictionmakes use of a generallyobserved
phenomenonin explorationprovinces.This phenomenon,whichmaybe termed"creaming", is
the diminishingeffectiveness of explorationeffortwithadvancingexploration.This pheno-
menon is so common that it characterizesall matureexplorationprovinces,and in fact,if
explorationis conductedwithany measureofefficiency, it is theprocesswhichall exploration
provincesmusteventuallygo through.
The creamingphenomenonis inpartan expressionofa declinein thesuccessrate,i.e.decline
in theprobabilityofsuccess.This in turnis a reflection ofthefactthatin a petroleumprovince
thereare a finitenumberofcommercialfields,so thatwitheach discoverythechanceofmaking
anotherdiscoveryis reduced.This decline,however,is to some extentcounteractedby the
industry'sincreasedcapabilityto locate new fieldsthroughadvances in technology.The net
resultis suchthattheprobability ofsuccess,evenin matureexplorationprovinces, declinesonly
gently,and its effect on the overallcreamingphenomenontendsto be relativelysmall.
A fargreaterrolein thecreamingphenomenonis thatofthefieldsize.Basicallythecreaming
phenomenonreflects diminishing fieldsize withadvancingexploration.This atteststo thefact
thattheindustry, throughitstechnologicalknow-how,is generallycapable offinding thelarger
fieldsearlyin exploration,and the smallerones in progressively laterstages.There may be
exceptionsto thisgeneraltrendwherebya "learningperiod",duringwhichthefieldsizetendsto
increase,may existin the earlystages of explorationin immatureprovinces.But once this
learningperiod is passed, the general decliningtrend is established almost invariably.
ExplorationtrendsofthetypeshowninFig. 1,characterizing matureexplorationprovinces, are
an expressionofthecreamingphenomenon,mostlyreflecting a declinein thefieldsize,and to a
lesserextenta declinein thesuccess rate.
The oil industryachieves the creamingphenomenonthroughsome sort of prospect
appraisalschemewherebyvariousexplorationprospectsin a provinceare rankedaccordingto
theirprospectivity as poor, promising,etc.,on the basis of available information about the
particularprospectsinvolvedand a wider experienceobtained fromother areas. As new
information (reviews,geophysicalsurveys,drillingresults)is gathered,the evaluationof the
prospectsis continuously updated.Allthisprocessresultsincharacterization ofeach individual
prospectin termsof (i) the probabilityof success,and (ii) the fieldsize, assumingdiscovery.
Assignmentofthesetwo parametersusuallyinvolvessubjectiveor judgementalprobabilities.
The estimationproceduremayrangefromveryinformal, unstructured well-
exercisesto formal,
structured decision-making processes.In a formalappraisalsystem,theprobabilityofsuccess
(sometimescalled thechancefactor)and thefieldsize estimatesmaybe combinedin a so-called
"expectationcurve" as shown in Fig. 2(a). [In factthe curve shows the so-called survival
function. Thus theterm"expectationcurve"is misleading, as itdoes notreferto expectationsin
thestatisticalsense.Nonetheless,thetermis widelyused withintheShellGroup,and we shall
use ithereas well.]In thisdiagrameach potentialfieldsize is plottedin appropriatevolumetric
unitsagainsttheprobabilitythatan actual fieldsize equal to or largerthanthepotentialvalue
willmaterialize.Fig. 2(b) showssome examplesofexpectationcurves.Such expectationcurves
can be used to rank explorationprospectsand help optimiseexplorationdrillingsequence
accordingto the oil company'sutilities.The creamingphenomenonobservedin exploration
provincesis essentiallythe resultof such rankingexercisesbased on technicalknow-how.
The creamingphenomenonfora realexplorationprovincecoded XX I 1 is illustrated in Fig.
3. Exceptfora scale difference along the y-axis,the explorationtrendshownwas the actual
observedtrendforthisprovincein theearly1970s.The generaldecliningtrendinthediscoveries
withadvancingexplorationis clearlyvisible.
1.4. GamblingAspect
Althoughadvancesin geologyand geophysicsmaysubstantiallyreducetheuncertaintyof
explorationdrillingoutcomes,explorationremainsverymucha gamble.A lotterybearsmuch

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
4 MEISNERAND DEMIRMEN- The CreamingMethod [Part 1,
PROBABILITY OF ACTUAL
FIELD SIZE BEING EQUAL TO EXPECTATION CURVE,
OR LARGER THAN POTENTIAL,0% PRINCIPLE
100

80

60

PROBABILITY
OF SUCCESS
40 -

Pi?
20 ;

0-
VI
POTENTIAL FIELD SIZE

PROBABILITY OF ACTUAL TYPES OF


FIELD SIZE BEING EQUAL TO
OR LARGER THAN POTENTIAL, % EXPECTATION CURVES
100 -

\ NO RISK, EXPECTATION
CURVEOF
\A DISCOVERYALREADYMADE
80 -

SMALL,BUT
6 RATHER CERTAIN
60

40 - \", EITHER/OR"TYPE SITUATION

20- VERY
POOR SPECULATIVE
PROSPECT
0 1
0 50 100 150
POTENTIAL FIELD SIZE, xIO6BBLS OIL

FIG. 2. Expectationcurves.

resemblanceto an explorationventure.The largenumberof structures or trappingconfigur-


ationsin a provincewhichmightcontainoil or gas are theticketsin thelottery,and thecostsof
concession,surveysand explorationdrillingconstitutethe price of a ticket.A prize is the
discoveryofa hydrocarbon accumulation,and theultimately recoverablevolume(thefieldsize)
is thereward,or thesizeoftheprize.Througha prospectappraisalsystem, ticketsbelievedto be
mostpromisingare boughtearlyin exploration,and those holdinglesserpromiseat a later
stage;unpromising ticketsare notboughtat all. The specificorderand stoppingruledependon
the particularsituationand the oil company'sutilities.Drillingan explorationwell means
finding out whetheror nota structure
containsoil and/orgas accumulationsand mayconsistof
not onlyone but severaldrillings.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNERAND DEMIRMEN- The CreamingMethod 5
CUMULATIVEDISCOVERY, BILLION BBLS
9

7?

6 _

2 -

PROVINCE XX iI
Number of wells 220
Cumulotivediscovery 7.00 blillionbbls
0*
0 50 100 150 200 250
NUMBER OF EXPLORATION WELLS
FIG. 3. Cumulativediscoveryagainstnumberof explorationwellsto illustratethe creamingphenomenon.

Our forecasting methodtakescognizanceofthegamblingaspectofpetroleumexploration.


It viewsthesearchforhydrocarbons as a stochasticprocesswithinherent risksand rewards,and
providesa forecastin termsof a predictiveprobabilitydistributionthatincorporatesall the
availableinformation fromthepastexplorationtrends.Inferences about crudeoil or naturalgas
discoveriesfromfutureexplorationactivitycan then be based on the entire predictive
distribution on the expectationcurvethatcorrespondsto thisdistribution.
or, alternatively,
2. THE DATA GENERATING PROCESS
Let us denotethe resultof the drillingof the nthexplorationwell by the randombinary
responsevariable x,nwherethe outcome x,n= 1 representsa discoveryor success and the
outcomexn,= 0 a dryhole or failure.The probabilityofsuccesswhenthenthexplorationwellis
drillingis denotedby On.Furthermore,in thecase ofa discovery, thefieldsize is denotedbythe
randomvariable vn.From experiencewe know that generallythe frequencydistributionof
reportedfieldsizesin a provincecan be adequatelyapproximatedby a lognormalprobability
distribution.Moreover,,the use of the lognormaldistributioncan also be justifiedon a
theoreticalbasis consideringthe formation,migrationand accumulationof hydrocarbon
deposits.So we shallassumethatthelogarithms ofthefieldsizes(discoveredand undiscovered)
withina provincefollowa normaldistribution withmean,a and variance U2. Throughoutthis
paper we will label a probabilitydensity function(p.d.f. for short) by its arguments.
Furthermore, ifIn Z is normallydistributedwithmean ji and variancea2, itS p.d.f.will be
denotedby No(InZ j,u,U2) and the p.d.f.of Z itselfby LNo(Z I ,, o2).

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
6 MEISNERAND DEMIRMEN - The CreamingMethod [Part 1,
Ifexplorationdrillingwerecompletelyrandomwithrespectto longitudeand latitude,then
theprobabilityofsuccesswouldbe approximately constantprovidedthatthesum oftheareal
extentsofthehydrocarbon accumulationwithintheprovinceis verylargebutrelatively smallas
comparedto thetotalarea ofthatprovince.In fact,theprobabilityofsuccessin randomdrilling
is roughlyequal to theratioofthesumofthearealextentsofall thehydrocarbonaccumulations
remainingto be discoveredto the total area of the provinceremainingto be drilled.
The areal extentofmanytypesofoil or gas reservoirs is highlycorrelatedwiththefieldsize.
So, evenin randomdrilling, largeoil or gas fieldshave a greaterchanceofbeingdiscoveredthan
small ones. Ifa discoveryis made in randomdrilling,theprobabilitythatthefieldsize willlie
betweenv and v+ dv willbe proportionalto vi wherethevalue ofexponentAdependson the
prevailingreservoir typein thatarea. In generalthep.d.f.ofthefieldsizein thecase ofdiscovery
is given by

p(vIu, c2,)= Vi p(V I , a2) V'i p(Vi,


I o2)dv, (2.1)

wherep(v I", U2) = LNo(v , 2). Aftersome algebraicreduction,the right-handside of (2.1)


turnsout to be equal to LNo(v | u+ c2, U2). Thus we may write
p(V Ii" U2, A) = p(V | 4, U2) = LNo(v | , U2), (2.2)
where4) = ji+ X'2.
Let us now consideragain scientific explorationdrillingthroughthe use of geologyand
geophysics.In binarydata analysis,a convenientwayto representa declinein theprobabilityof
successwithadvancingexploration,so thattheconstraint < On<1 is inevitablysatisfied, is to
postulatethe linearlogisticmodel
on
=
{1+exp(oc1+ 0C2n)} 1, C2 0, (2.3)
wherethecululativenumberofexplorationwellsnis usedas a measureofexploratory To
effort.
represent a declinein thefieldsize withadvancingexploration,we postulatethattheexponentA
in (2.1) is notconstantbuta lineardecreasingfunction ofthecumulativenumberofexploration
wells,thus
An = 71+72n, 72<0. (2.4)

As a result,themeanOn ofthelogarithmofthefieldsize vnin thecase whenthenthexploration


well turnsout to be a successcan be writtenas
= j+y1C22 +
,n, 2 n=f31 +f32n, f32-< (2.5)
where f31= ji+y1 U2 and fB2 = )2 U2. So, the p.d.f.of Vn iS givenby

P(Vn If31+ 32 n, U2) = LNo(vn If31+2 n, U2). (2.6)


the
Let Rnrepresent totalultimately recoverable reservefrom discovered
fields by n exploration
wellsand rnthecontributionof thenthexplorationwell to thistotal reserve.Hence we have
rn= xn vn (2.7)
and
n

n= ri. (2.8)
i= 1

Thus it is assumed that the observedsequence of drillingresultsconstitutesa random


observationofthesequenceofindependentrandomvariablesr1,r2,...,rngeneratedbythedata
generatingprocess definedabove. From the propertiesof the lognormaland the mixed

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNERAND DEMIRMEN- Method
The Creaming 7
lognormaldistribution(see Aitchisonand Brown, 1963) we have for the expectationand
varianceof Rn:
n

E(Rn)= E 01 i (2.9)
i= 1
and
n

var(Rn) = E OEi7(0)-O), (2.10)


i= 1
where
Xi = exp(f31+f2 i+o+ ) (2.11)
and
w = exp U2. (2.12)
Furthermore,thecumulativedistribution
function(c.d.f.forshort)oftheadditionalreserveri
froman explorationwell withserialnumberi is completelyspecifiedby the parameterset
4i = {i,gX1,CX2,f ,1 2,a2} (2.13)
and is givenby
0
O forr1<0,
P(rij )= (2.14)
f(1 OiO+
O LNo(rI1T +#2i,G2)dr forr O.
f
The c.d.f.ofRnfollowsfromtheconvolutionofthec.d.f.s.ofr1,r2,...,rnbutcannotbe obtainedin
a mathematically tractableform.
3. DATAANALYSIS
An analysisofdata ofpast explorationdrillingperformance in a provincemustprovideus
withthebuildingstonesforpredictionoffutureperformance. The data requiredconsistofthe
results of the exploration wells drilled. They should be listed in historical sequence,
distinguishingdiscoveries(even if non-commerical) fromdryholes. For the discoveriesthe
amountofultimaterecoverableoil or gas shouldbe given.The data thusassembledand "rolled
back" to thetimeofdiscoverywillidentify thepast explorationtrends.Giventhelinearlogistic
model,the likelihoodof the sequence of dryholes and discoveriesis givenby

Y ix)}
H O~(l~~O~) exp{-Q?1 E xi+c02
L(Xn Ia) = rI OiN
(-oi O '=
x= 1 = (3.1)
i-
1 IH [1 +exp {- (1+02 i)}]
i= 1
wherexnis thevector(x1, x2,..., xX)' and a = (och,
02). In matureexplorationareas we wantto
make our forecastson the basis of the data only,so we use neutralpriordistributions with
constantdensityforoc,and 02. As a resultthe posteriordistributionof a, apart fromthe
normalizingconstant,has the p.d.f.

P(E IXn)oc
exp{-(t 1 1l+ t1 2 2)} (3.2)
Hl [1 +exp{-(e1j+ 02 i)}]
i= 1

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
8 MEISNERAND DEMIRMEN - The CreamingMethod [Part 1,
The quantitiest1=1 xi and t2 = 1% 1ixi are minimalsufficientstatisticsfor witho2
fixedand 062 withoc,fixed,respectively. oc,
Hence theycontainbetweenthemall theinformation
inherentin thesequenceofdiscoveriesand dryholes.So, theposteriordistribution is completely
specifiedby thesesufficientstatisticsand n, the numberof explorationwellsdrilled:
p(a I Xn)_ pa In,t,1,tl 2). (3.3)
Althoughp(a In,t,1,t12) can be readilyevaluatednumerically, itis,forlargen,moreconvenient
to approximateit by a bivariatenormal distributionwith means equal to the maximum
likelihoodestimatesoc1, and &c22,and a dispersionmatrixD(a) whose inversehas as elements
minusthesecondpartialderivatives ofthelogarithmofthelikelihoodfunction (3.1)withrespect
to theparametersoc,and 062 takenat its maximum.The validityoftheappoximationfollows
fromthe largesamplepropertiesof the Bayesianposteriorp.d.f.'s.Thus,
p(|n t , tl2) No[a| &,D(a)], (3.4)
whereNo[a I&,D(a)] denotesthe bivariatenormalp.d.f.withmean vector& and dispersion
matrixD(a).
The logarithmof the likelihoodis givenby
n

InL =(tl 1 01 + t12 062)- ln [I + expI{- (ot + 02 i)}] (3.5)


i= 1

and the maximumlikelihoodestimateso&,and &c2 followfromthe equations:


n
= -t1l+
alnL/aoc1 {1+exp(x1 +02 i)} =0

alnL/a062
= -t12+
n
{1 +exp(x1 +02 i)}
._
=0 I (3.6)

The inverseof thedispersionmatrixD(a) is givenby


/n n\

E OiG - Od E ioi( - 0i)


= 1 ~i= 1 (A1
D-l(a) i-l
n i-l , whereOi= {1 +exp( o+o2i)}'. (3.7)
, ioi(l-i) 2il0
\i=l i=1 /
The likelihoodequations(3.6) are readilysolved by a Newton-Raphsonprocedure.
oftheparametervectorP = (fll, /2)1 and thevariancea2 inour
The posteriordistribution
lognormallinearmodel (2.6) followsdirectlyfromnormalregressiontheory(see Raiffaand
Schlaifer,1961).Firstwe summarizethedata in the followingset of sufficient
statistics

T= (t1 t12) (3.8)

wheret22 == 1 x, and
/n n

g= Y xi ln vi,E ixiln vi (withxiln vi-0 forxi = 0). (3.9)


i=l i=l
From Tand g we obtain
T=Tg- (3.10)

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNERAND DEMIRMEN- The CreamingMethod 9
Furthermore,
v = t11-2 (redundant) (3.11)
and

s= { (xiInvi)2-1 T i}v (3.12)

in termsof the precisionconstanth = 1/a2 as


Then,writingthe posteriordistribution
p(p,hIp ,7S2, v) =p(p lP, T,h)p(hlIS2,v), (3.13)
priorsforJand h,
we have, usingnon-informative
p(,B T,7 h) = No[P |r(hT)- 1] (3.14)
and
p(hIS2, V) = Ga[hIS2, V] (3.15)
whereGa[h js2, v] denotesthe gamma p.d.f.withparameterss2 and v. So all the information
containedin thedata concerningtheparametersin our data generatingprocessis completely
summarizedby
y = {n,X g,s2, v}.
We shall use theexplorationdrillingperformance data shownin Fig. 3 fromtheprovince
XXII to illustratethedata analysis.In Table 1 thefieldsizesofthesuccessfulexplorationwells
are listedtogetherwiththeirserialnumberofdrilling.The data fromthefirst180 exploration
wellswillbe usedforour data analysis.Subsequentlywe willforecastthetotalreservefromthe
remaining 40 wellson thebasis oftheresultsofthisanalysis.The forecastcan thenbe compared
withthe actual figureso as to judge theircorrespondence.

TABLE 1
Field sizes of58 discoveriesout of220 explorationdrillings

Exploration Field size Exploration Field size Exploration Field size


wellNo. oil 106 BBLS wellNo. oil 106 BBLS wellNo. oil 106 BBLS

3 28 56 177 141 88
7 26 57 43 144 29
8 775 58 33 145 450
9 114 62 178 152 59
11 31 71 15 161 88
12 337 76 22 162 49
17 41 78 11 171 100
18 113 81 81 174 10
20 1328 82 35 176 88
21 21 88 25 178 17
22 13 90 170 -
29 455 92 19 189 12
33 89 102 56 195 125
34 482 105 42 197 20
35 70 109 335 202 88
39 215 110 21 203 88
45 62 116 50 209 69
46 58 119 181 210 25
52 69 131 93 215 100
55 154 139 75

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
10 MEISNER AND DEMIRMEN- The Creaming Method [Part 1,

containedin the data fromthe first180 explorationwellsin Table 1 is


The information
completelysummarizedby
n= 180,
(50 3942
3942 454 792
= (4-643,-8&465 x 10- 3

s= 1624 withv = 48.


By solvingthe maximumlikelihoodequations(3.6) we obtain
= (0-4277,6-075x
7 0.10285 - 8&956x 10-4
D(at)=
8956x 10-4 1 0752x 10i- /
Fig. 4 showsthe90 percentHighestDensityRegion(HDR) fora. Thisfigureshowsalso forthe
midpointsina gridcoveringthis90 percentHDR theactualposteriorp.d.f.computedfrom(3.2)
0C2

6.9 11.6 5.3 0.7 - UPPER FIGURE: SCALED ACTUAL DENSITY


SCALED DENSITY OF
7.27.2 10 .10. 3.6
.6* 00. 3 ---LO LOWER
WE R FlGsURE:
IGURE:NORMAL APPROXIMATION
0.012 -
4.8 24.6 30.6 10.2 1.0

5.4 25.4 28.7 7 0.5


0.0,10
0.8 15.0 58.4 54.4 i13.5 1.0
59.3 53.0 0.6
0.008 - 1.1 1 11.5

0.0 2.2 30.2 91.0 65.1 12.5 0.7

0.1 2.6 3.6 91.8 64.8 11.1\, 0.5


0.006 -
0.1 3.5 39.3 93.4 52.9 .1 0.4

0.0040.1 4.4 41.1 94.3 52.5 1 0.2


0.004 -_ _ _

0. 1 3.6 32.8 63.4 29.4\ 3.7

0.2 4.7 35.4 64.1 28.2 3.0


0.002 - \
0.1 2 17.6 28.6 12
0 0
~~~~~~0
0.1 3.4 0. 28.9 10.0

0.0 0.9 6
0.1 1.6 7.6 8.6
-0.002.1 1
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1.0 1.2

0( i
FIG.4. Ninetypercenthighestdensityregionforot and (2 comparisonofactualand approximatedposteriordensitiesof
ociand 2

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] - The CreamingMethod
MEISNERAND DEMIRMEN II
and thenormalapproximationaccordingto (3.4).The densitiesshownare scaledin sucha way
thatthedensityofthenormalapproximationat themaximumlikelihoodestimatesequals to
100. It can be seen that the normal approximationis quite adequate. From the marginal
posteriordistribution of x2 we inferthattheodds in favourof a declinein theprobabilityof
successwithadvanced explorationare roughly30 to 1.
In Fig. 5 we have plottedthe observedprobabilityof success in seriesof 10 successive
explorationdrillingsagainstthemeanexplorationwellnumberin each series.Thisfigureshows
also thelinearlogisticformfittedto thebinarydata so as to showthecorrespondence.
The data

PROBABILITY OF SUCCESS OBSERVED PROBABILITY DATAUSED IN


OF SUCCESS IN SERIES J PARAMETER ESTIMATION
.6 OF 1O SUCCESSIVE
EXPLORATION DRILLINGS 0 TEST DATA

.5

.4

.2

O- I E
I , , , '

0 50 100 150 200 250


NUMBER OF EXPLORATION WELLS

FIG. 5. Linear logisticdeclineof the probabilityof success withadvancingexplorationin provinceXX11.

analysisin our forecasting procedureincludesan analysisofresidualsalong thelinessuggested


byCox (1970).In thisanalysiswe approximatethedistribution ofthenumberofdiscoveriesin a
seriesof 10 successiveexplorationdrillingsby a binomialone witha probabilityofsuccessas
predictedbythelinearlogisticformfromthemeanexplorationwellnumberin thatseries.The
difference betweenactual and predictednumberofdiscoveriesin each seriesis thendividedby
itsapproximatestandarderrorand thestandardizedresidualsthusobtainedareplottedagainst
themeanexplorationwellnumber.This plotis used to examinetheadequacy offitby looking
forsystematic departuresfroman expectedrandompattern.Furthermore, wecomputetheratio
of the likelihoodswith probabilityof success as observedin each series of 10 successive
explorationdrillingsand thatpredictedby thelinearlogisticform,respectively. Two timesthe
logarithmofthisratiohas approximately a x2distributionwitha numberofdegreesoffreedom
equal to thenumberofseriesminustwo.So, thisquantitydividedby itsnumberofdegreesof
freedomwill have an expectationequal to one, and may serveas a yardstickto assess the
goodness of fit.In our examples we could not trace in the plot of standarizedresiduals
systematic departuresfroma randompattern.The degreeoflack offitwas equal to 1 14,which,
givenitsdistribution, is quiteclose to itsexpectation.So, theresultsare quiteconsistentwith
whatcould be expectedon the basis of the assumed binarydata generatingprocess.
Fig. 6 showsthefieldsizeplottedagainsttheexplorationwellnumberand thefitted loglinear
regression line.Fromthemarginalposteriordistribution oftheslope/2 weinferthattheodds in
favourofa declinein thefieldsize withadvancingexplorationare roughly150to 1. In orderto
checkthe adequacy of our loglinearnormalmodel (2.6) the data analysisin our forecasting

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
12 MEISNERAND DEMIRMEN- The CreamingMethod [Part 1,
FIELD SIZE, OIL 06 BBLS DATA USED IN LOG/LINEAR
REGRESSION ANALYSIS
2000
0 TEST DATA
1000

500

200-

1oo

50-

20-

10- OD

5-

2-

0 50 100 150 200 250


NUMBER OF EXPLORATION WELLS
FIG. 6. Loglineardeclineof the fieldsize withadvancingexplorationin provinceXX11.

procedureincludesan analysisoftheso-calledrecursive residualsalongthelinesas suggestedby


Brownand Durbin(1968).Let jk- 1be thecoefficient vectorofthefitted line
loglinearregression
on thebasis ofthefirst k - 1 discoveriesand letnk - 1 referto theserialnumberoftheexploration
drillingthat yielded the (k- 1)th discovery.Accordingto normal regressiontheorythe
redictive distribution ofthelogarithm ofthefieldsizeofthenextdiscoveryis normalwithmean
A1k- 1 + f2,k - 1 nk and variance 4kU2 where 4kis givenby
(k-1 -1
Ak = 1 +(k-1) 1+(nk-n-k-1
)2i (ni -nk-_1)29 (3.16)

and -nk-1 = Eki 1 ni(k - 1).The recursive


residualsWk, k = 3,4,..., t11,aredefined
as
wk = (ln Vnk
-, /k-1 - 2,k-1 nk)l/\/k- (3.17)
Consequently, givenourlognormalmodel,thewk'sareindependently with
normallydistributed
zero mean and variancea2. In theanalysisoftheserecursiveresidualswe applythefollowing
techniques:
(1) A normalplot of the recursiveresidualstogetherwiththe computationof Shapiro and
Wilk's W (1965) if the numberof discoveriesis less than or equal to 50 or otherwise
d'Agostino's D (1971), in order to assess the consistencywith the assumption of
lognormality.
(2) The computationof von Neumann'sratio as suggestedby Phillipsand Harvey(1974) in
orderto assess the consistencywiththe assumptionof mutualindependence.
(3) A plot of the cusum quantityof the standardizedrecursiveresiduals
k

Y wk/s, k =3,4...,tll
i=3

againstk,togetherwithappropriatecontrollimits,in orderto assess whethertheloglinear


relationshipis constantover the explorationplay.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNER AND DEMIRMEN - The CreamingMethod 13
(4) A plot of the cusum quantityof the standardizedsquare recursiveresiduals
k

E Wk2/VS2, k = 3,4,..., t
i= 3

againstk togetherwithappropriatecontrollimits,in orderto assess whethervariancec2 iS


constantover theexplorationplay.
Figs 7, 8 and 9 showtheresultsofthisanalysisofrecursiveresidualsforthefieldsizedata of
the50 discoveriesin provinceXX11.The normalplotin Fig. 7 does notshowseriousdepartures
fromthe straightline,whichrepresentsa normaldistributionwithzero mean and variance

01
0.2

0.5 -

2 -

5-

so -

/
20 - <

30 0

40 - 0
40
50 -
70 -~~~~~~~

80 1
90 -

95-

98 - s = 1.274
99 SHAPIRO AND WILK'S W = 0.97
99.5 - VON NEUMANN'S RATIO = 2. i i

99.8 -

99.9 _ _15F- I I I I I I
-3 -2 -4 0 4 2 3
RECURSIVE RESIDUAL

Fig. 7. Normal plot of the recursiveresiduals.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
14 - The CreamingMethod
MEISNERANDDEMIRMEN [Part 1,
CUSUM OF STANDARDIZED
RECURSIVE RESIDUALS

+ 16 -

+ 14 -

+ 12
90 % UPPER CONTROL LIMIT

+ 10

+8a

+ 6

+ 4

+ 2

-2

-4

-6

-8

- 10-\

90% LOWER CONTROL LIMIT


- 12-\

- 14
0 1O 20 30 40 50
NUMBER OF DISCOVERIES
FIG. 8. Cusum plot of standardizedrecursiveresiduals.

quite close to its


equal to S2. Moreover,Shapiro and Wilk's W is, given its distribution,
expectation.The same holdsforthevonNeumannratio.The 90 percentcontrollimitsin Figs 8
and9 aredrawninsuchaway thattheprobabilityofthesamplepathcrossingone or bothlinesis

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNERAND DEMIRMEN- The CreamingMethod 15
approximately10 per cent.The slopes ofthecontrollimitsforthecusumof thestandardized
recursive thata singlepointofthesample
residualsin Fig. 8 arechosensuchthattheprobability
path lies outside the limitsis maximumhalf-waybetweenk = 3 and k = t11. The proper
functionof the controllimitsis merelyto provide yardsticksagainst which to assess the
observedpatternsofthesamplepath.It can be seen fromFigs 8 and 9 thatbothsamplepaths

/
CUSUM OF STANDARDIZED
SQUARED RECURSIVE RESIDUALS
/

1.0 //

0.9/

0.7 /\ 4/

0.8 /
0.4-

0
11 20 30 40 50
0

NUMBER OF DISCOVERIES

FIG. 9. Cusum plot of standardizedsquared recursiveresiduals.

look well-behaved.So, in this example the data behave quite consistentlywith the posed
lognormallineardata generatingprocess.
The analysisof residualscompletesthedata analysis.In thenextsectionwe willdescribe
ofthetotal
how theresultsof thedata analysisare used to obtain thepredictivedistribution
ultimatelyrecoverablereservefromfuturediscoveries.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
16 MEISNERAND DEMIRMEN- The CreamingMethod [Part 1,
4. FORECASTING
Let m representthenumberof additionalexplorationwellsyetto be drilledand Rnm the
total ultimatelyrecoverablereservefromthefuturediscoveries.Hence we have
m
Rn,m E rn+j- (4.1)
j= 1

Accordingto (2.14) thec.d.f.oftheadditionalreservern+lfromthe(n+j)th explorationwellis


givenby
O forrn+j<O
P(rn+~j4*n+j) = (4.2)
= I On+j+ On+j
(rn+i

I +/
LNo[z
/3 2(n+j), U2]dz forrn+j>,O

where
n+j = [1 +exp {c1 + OC2(n
+j)}] - (4.3)
and
pn+j = {n+j,a, , a2}. (4.4)
as a resultofourdata analysistherespective
Furthermore, posteriordensitiesofa, f and U2 are
givenby
p(a In,t 1, t12)= No[a &,D(a)], (4.5)
p(p T,7h) = No[jJ ,(hT)-1] (4.6)
and
p(hIS2, v) = Ga[hIS2, v), (4.7)
whereh = 1/U2.
Theoretically, thederivationof thepredictivec.d.f.of Rn,mis a straightforward
processin
whichno particulardifficulties are met.It consistsof the followingtwo steps:
(1) The c.d.f.ofRn,mgiventheparameterset {n, m,a, j, 2} followsfromtheconvolutionofthe
a

c.d.f.sof rn+j j = 1,2,...,m. The resultingc.d.f.of Rn, will be the mixture

0 forRn m< 0

P(Rn,
ml M,ot~P, a2) _i m
p(zlI1,
P,h)dz forR (4.8)
1=2o

whereil, 1= 1,2,...,2mare thesuccessivetermsin theexpansionof171= J{(1 - On+j)+0


I
and the p.d.f.p(z j1,j, h ') is the convolutionof the p.d.f.'sLNo[vn+j I3l+# 2(n +j), h- ]
withindicesn+j correspondingto the factorsOn+jappearingin tl.
(2) The q,, 1= 1,2,'...,2, are multipliedby the posteriorp.d.f.No[a I&,D(a)] and the p.d.f.s
p(z I 1,f,h- 1),1= 2,3,...,2mbytheposteriorp.d.f.No[p IP(hT)- 1] Ga[h Is2, v]. Subsequently
a, P and h are integratedout.
However,theresulting predictivec.d.f.P(Rn,m m,y)cannotbe obtainedin analytically tractable
closedform.Therefore we use a MonteCarlo procedurein practicalapplications.An outlineof
this Monte Carlo procedureis givenbelow.

Monte Carlo procedureto evaluatethepredictivec.df of Rn, m


Start
(1) SpecifythenumberN of simulationruns.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNER AND DEMIRMEN- The Creaming Method 17

(2) Specifythenumbern ofexplorationwellsalreadydrilledand thenumberm ofexploration


wells remainingto be drilled.
(3) Specify{&,D(a)}, {p,T} and {s2,v}.
(4) For i = 1,2, ..., N:
(4.1) Generatea parametervectora fromthe posteriorp.d.f.No[a Ia, D(a)].
(4.2) Generatea precisionconstanth fromthe posteriorp.d.f.Ga[h Ls2,v].
(4.3) Generatea parametervectorP fromthe posteriorp.d.f.No[P I ,(hTT)].
(4.4) Set Rn,mequal to zero.
(4.5) For j]1, 2, ..., m:
(4.5.1) Compute On = [1 +exp {c1++C2(n +j)}]
(4.5.2) Generatea value xn+joftherandombinaryresponsevariablewithprobability
of successequal to On+j
(4.5.3) If xn+j = 0 thenrn+j= 0 and jump to (4.5.6),otherwisecontinue.
(4.5.4) Generatea fieldsize vn+ifromthe p.d.f.LNo[vn+j 31+/32(n+j), h'].
(4.5.5) rn+j=Vn+j
(4.5.6) Rn,m= Rn,m+rn+j.
(4.5.7) Returnto (4.5.1).
(4.6) Save Rn,mand the numberof discoveries.
(4.7) Returnto (4.1).
(5) Group theN values ofRn,m thusobtainedand thenumberofdiscoveriesintoa frequency
distribution.
Stop.
Figs 10, 11 and 12 show the resultof the forecastingprocedurefor the 40 additional
explorationwellsinprovinceXXII on thebasis oftheresultsofourdata analysis.Fig. 10 shows
theexpectationcurveofthetotalultimaterecovery, Fig. 11 thep.d.f.ofthetotalrecoveryfrom
discoveriesand Fig. 12 the probabilitymass functionof the numberof discoveries.In this
example2000 simulationrunswereused.The expectationcurvein Fig. 10 incorporatesall the
available informationfrom the past exploration trends. Relevant inferencesabout the
prospectiverecoveryfromthe40 additionalexplorationwellsshouldbe based on thewholeof
thisexpectationcurve.
Sometimesa simplified oftheexpectationcurvemaybe neededin termsofa
representation
fewnumericalvalues.Accordingto a proposednomenclaturein Shell,theexpectationcurveis
characterizedby means of the followingfivevalues:
(1) The chance factorc, whichis the probabilityof at least one discovery,in our example
c = 0993.
(2) A low value L corresponding to a probabilityof lc thattheactual ultimaterecoverywillbe
less thanL, in our exampleL= 75 x 106 barrels(BBLS).
(3) A middlevalueM corresponding to a probabilityof4c thattheactualrecoverywillbe either
more or less than M, in our exampleM = 227 x 106 BBLS.
(4) A highvalue H corresponding to a probabilityof lc thattheactual recoverywillbe more
than H, in our exampleH = 546 x 106 BBLS.
(5) The "expectation"of the ultimaterecoverydefinedas thechance factormultipliedby the
arithemeticmean of L, M and H, in our example the expectationof the ultimate
recovery= 281 x 106 BBLS.
It mustbe emphasizedthatthesefivevaluesmerelyserveas a simplified ofthe
representation
expectationcurve.Theirinterpretation in the
is directlyrelatedto thepredictivedistribution
formoftheexpectationcurve.The expectationoftheultimaterecoverymaybe regardedas a
pointestimatefordescriptivepurposes.
To concludethissectionwe notethatin provinceXXII the40 additionalexplorationwells
resulted in eight actual discoveries, with a total (additional) ultimate recovery of
305 5 x 106 BBLS.These resultsare well in line withour forecast.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
18 MEISNERAND DEMIRMEN- The CreamingMethod [Part 1,
PROBABILITY IN % THAT
ACTUAL RECOVERY IS EQUAL TO
OR LARGER THAN POTENTIAL

99.8
99.5
99 v

98-

90 \

80 \
70-
60 \
50 -

30 \
20 -

10 \

5-

0.5
0.2
0.1 ACTUALRESERVE FROM
THE 40 ADDITIONAL
EXPLORATION WELLS

f yi I I , I I I I, I- - I
0 0.2 0.5 1 2 5 10 20 50 100 200 500 1000 2000 5000
POTENTIAL RECOVERY, OIL 106 BBLS
FIG. 10. Expectationcurveof ultimately
recoverablereservesof discoveriesfrom40 additonal explorationwells in
provinceXX11.

5. DISCUSSION AND CONCLUDING REMARKS


In theprecedingsectionswe havepresenteda procedureto forecastcrudeoil and naturalgas
reservesfromfuture discoverieson thebasis ofthecreamingphenonemongenerallyobservedin
maturepetroleumprovinces.Therefore we call our methodthe"creaming"method.Although
to our knowledgethecreamingmethodis newin itsapproach,forecasting procedureswiththe
same objectiveshave been used in thepast.A numberofworkers,among themMoore (1962),
Zapp (1962), Hubbert(1967),Arps et al. (1971), Pelto (1973), Richards(1973) and Odell and
Rosing (1974), have devised or utilized statisticalor mathematicalmodels to forecast
hydrocarbondiscoveries.These workersconsideredtime or cumulativebore-hole length
("footage")as an indicatorforforecasting whereaswe employthenumberofexplorationwells
plannedto be drilledin thefutureas an indicator.In ourviewthenumberofexplorationwellsis
a betterindicatorthantimeas thelatteronlyindirectly controlsfuturediscoveries,themore
directcontrolbeingtheexploratoryeffort expended.In thisrespectfootageas an indicatoris
nearlyas good as thenumberofexplorationwellsbutfootagehas thedisadvantagethatitdoes

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNER AND DEMIRMEN - The CreamingMethod 19
PROBABILITY DENSITY x10S

300

250 -

PROBABILITY OF NO DISCOVERIES = 0.007


200-

I: I\

150 I I
\ \

I I \
150 I I
so I I

I I I-
100 I .

50
ILOW II MIDDLE HH

0 500 1000 11500


TOTAL RECOVERY, OIL 106 BBLS
recoverablereservesof discoveriesfrom40 additional
FIG. 11. Predictiveprobabilitydensityfunctionof ultimately
explorationwellsin provinceXX1.

notallowa separateanalysisoftheprobability ofsuccessand thefieldsize,as in ourapproach.A


methodrecentlysuggestedby van Eek (1974) makes use of thecumulativeproductionas an
indicator.The appropriatenessof thisindicatoris also open to questionsinceincreasedpro-
duction,exceptperhapsin theshorttermand inspecialcircumstances, can hardlybe considered
to lead to morediscoveries;infact,providedthatdemandand economicincentives are there,the
oppositecan be said to be true.None ofthesemethodsbyearlierworkersyieldsforecastin the
formofa predictivedistribution and none,apartfromtheone proposedby Pelto (1973),gives
forecaststhatare interpretable in a probabilitycontext.Also noteworthy amongtheprevious
worksare theempiricaland theoreticalstudiesconductedby Arpsand Roberts(1958) and by
Kaufman(1965) on theprobability offieldsizes.Ryan(1973)gave a modelrelating
distribution
discoveryrate to comulativenumberof wellsin Alberta,but his model,intendedmainlyfor
describingpast discoveryrates,is notreadilysuitableformakingoil or gas discoveryforecasts.
Finally,we only recentlybecame aware of a paper by Kaufman et al. (1975) in whicha
probabilisticmodel of oil and gas discoveryis also described.They compared statistical
properties ofmajorAlbertaexplorationplayswithproperties ofa MonteCarlo simulation.This
model is quite differentfromours and has not yetbeen empiricallyvalidated.
The data analysisto obtaintheposteriorp.d.f.'softheparametersoccurringin our modelis
straightforward and applies well-knowntechniquesof maximumlikelihoodestimationof
parametersin the linearlogisticformand lognormallinearregressionanalysis.Since all the
information frompast explorationtrendscan be completelysummarizedby a set ofsufficient
statistics,updatingtheposteriorp.d.f.'softheparameterswhennewdata becameavailableis a

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
20 MEISNER AND DEMIRMEN - The CreamingMethod [Part 1,
PROBABILITY, %
15 1

ACTUAL NUMBER OF DISCOVERIES - 8

10 | MODE = 6
MEDIAN= 6
MEAN =6.64

0 l 2 3 4 5 6 7 8 9 10 112 13 14 15 16 17 18 19
NUMBER OF DISCOVERIES

FIG. 12. Predictiveprobabilitymass functionof the numberof discoveriesfrom40 additionalexplorationwells in


provinceXX1.

rathersimple process (see Raiffaand Schlaifer,1961; Walker and Duncan, 1967). In each
applicationthe adequacy of our model to describeproperlythe past explorationtrendsis
assessedon thebasis ofan analysisofresiduals.The teststatistics, togetherwiththeirsample
distributions,are merelyused as yardsticksagainst whichthe questionablediscrepanciesas
observedinthevariousplotsareassessed.As a resultoftheanalysisofresidualswe foundthatin
some provincesthe probabilityof success and/orthe fieldsize were constantor grewto a
maximumbeforetheystarteddeclining,thusindicatinga "learningperiod".In orderto cope
withthiskindofdeparturefromour modelwe triedtwodifferent remedies:(1) to reparametrize
our model to allow forthelearningperiodand (2) to regardthelearningperiodas irrelevant
information and restrict
ourdata analysisto theexplorationtrendafterthelearningperiod.The
pointwhereto startcan be readilytracedbyrunningthecusumplotsofstandardizedrecursive
residualsbackwardsthroughtheexplorationplayinsteadofforwards. The secondapproachof
neglectingthelearningperiodyieldedthebetterresults.Sometimeswe foundthat(apartfroma
possibleinclinein theearlystage)past trendsofboth probabilityofsuccessand fieldsize are
constant.Thentheprovinceis in a transition stageofexplorationfromimmatureto mature.In
thiscase the creamingmethodcan stillbe used to make a short-term forecastof the total
recoveryfromexplorationwellsto be drilledin the near future, say a year.
In orderto havesomecheckon theresultsoftheMonteCarlo procedurewecomputealso an
estimateof the mean of the sample distribution of the total ultimaterecoveryRn,m givenby

exp{fll1+f2(n+j) +1i2}
E(Rn,m)M
J1 1+exp {x1+oc2(n+J)}j (5.1)
To estimatethe numeratorexp{/lh
+,/2(n+j) +, j2} we make use of the unbiasedminimum
varianceFinney-type
estimatorsforlognormallinearmodelsderivedby Bradu and Mundlak

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] MEISNER AND DEMIRMEN - The CreamingMethod 21
(1970). In our applicationswe observeda close agreementbetweenthe pointestimateof the
sampledistribution thuscomputedand themedianof thepredictivedistribution of Rn m. We
mustemphasizethatthiscomputedpointestimateis used onlyas a check.We do not see much
use in computingpointestimateswithspecificsamplingproperties, suchas unbiasedminimum
variance,minimummean squarederroror optimalrelativeto a squared errorloss function, if
theyexist at all (in the convolutionprocess of derivingthe predictivedistributionof Rn m
logstudentdistributions willappear whichhave infinite mean and variance).Inferences about
thetotalrecoveryRn m shouldbe based on theentirepredictivedistribution whichcontainsall
theinformation availablefromthepastexplorationtrends.In a realformaldecisionanalysisin
whichfuture explorationeffort is weighedagainstpotentialgains,theoil company'sutilitiesare
determinedby manyfactorsand cannot be approximatedby relativelysimplemathematical
loss functions.
Thereis a world-wide tendencyto underestimate thefieldsizesofrecently madediscoveries,
whichwill deflatethe forecasts.Thereforeour procedureshould make the best allowances
possible forappreciationin fieldsizes withtime,due to such causes as fieldextensionsand
improvedrecoverytechniques,theso-calledrevisions.If we make thesimpleassumptionthat,
on average,therevisionswillenlargethefieldsizesin a provincebya constantfactor,a solution
forthisproblemis readilyavailable. Systematicunder-or overestimation ofthefieldsizes in a
provincebya factorA willonlyaffect theintercept ,Bloftheloglinearregression in thefollowing
way#1 revised= pl1 + lna. So, a newexpectationcurvecorrectedfortherevisionscan thenbe
readilyderived.
Finally,we considerit worthwhile to add a cautionaryremarkregardingtheapplicationof
our method.Apartfromtherequirement thattheprovinceforwhicha forecastis made should
be in a matureexplorationstage(or at leastin a transitional stageforshort-term forecasts),itis
also necessarythat the particulararea under considerationwas, at one time or another,
available or accessiblein its entiretyforexploration.We stressthe words "available" and
"accessible",forthereis no requirement thatthearea in itsentirety was actuallyexplored,only
thatthepetroleumindustry, or otheragenciesactivelysearchingforoil or gas,iftheychose,were
in a position,technologically or in termsofpermission, to do so. This is a commonlimitationof
almost all statisticalforecastproceduresthat try to avoid "surprises"or "freaks",but
nonethelesswe considerit wortha mentionas it is a limitationeasy to overlookand, ifnot
heeded,maylead to misinterpretation offorecastresults.In ourcase,thelimitation followsfrom
our use ofthecreamingphenomenonas an underlying conceptfortheforecasting procedure.
Basically,the limitationmeans thatin makingforecastswe should be acutelyaware of the
particulargeographicarea or geologicdomainforwhichwe are makingforecasts, byexcluding
fromconsiderationthosepartsofthegeographicarea or geologicdomain thatwerenot,and
could not be,exploredbefore.Hence,forexample,deep sea prospectsshouldneverbe lumped
with land prospectsin making a forecast,and similarlyprospectslocated in a previous
concessionarea should be kept separatefromprospectsin an area that was neveropen to
exploration.Deep prospectsin an area,however,could be lumpedwiththeshallowprospectsif
theindustryin generalhad thetechnologicalcapabilityto drillthedeep prospectsin thepast.
We have applied thecreamingmethodto a selectednumberofpetroleumprovincesin the
worldto obtainindustry-wide oil discoveryforecastsfortheperiodsince 1973.The potential
and robustnessof the methodwas also testedthrougha seriesof simulatedforecastsusing
controlledsituations,and practicaland theoreticalramifications wereinvestigated. The results
indicatethat the methodshould have a wide applicationformakingoil and gas discovery
forecastsforplanningpurposes.

ACKNOWLEDGEMENTS
We aregratefulto themanagementofShellInternationale PetroleumMij. B.V.(The Hague)
forpermissionto publishthispaper.Thanksare also due to our manycolleaguesin theShell

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
22 Discussionof thePaper by Mr Meisnerand Mr Demirmen [Part 1,
organizationfordiscussions;in particular,Dr M. H. Nederlof(SIPM) offeredmuch useful
advice and also read the manuscript.
REFERENCES
AITCHISON, J. and BROWN,J.A. C. (1963). The LognormalDistribution. Cambridge:UniversityPress.
ARPS,J. J., MORTADA, M. and SMITH A. E. (1971). Relationshipbetweenproved reservesand exploratoryeffort.
J. PetroleumTechn.,671-675.
ARPS,J.J.and ROBERTS, T. G. (1958).Economicsofdrillingforcretaceousoil on east flankofDenver-Julesburgbasin.
Amer.Assoc. PetroleumGeologistsBull.,42, 2549-2566.
BRADU, D. and MUNDLAK, Y. (1970). Estimationin lognormallinearmodels.J. Amer.Statist.Ass.,65, 198-211.
BROWN, R. L. and DURBIN, J.(1968). Methodsforinvestigating whethera regression relationshipis constantovertime.
Amsterdam:(SelectedStatisticalPapers fromthe European Meeting)MathematicalCentreTractsNo. 26.
Cox, D. R. (1970). AnalysisofBinaryData. London: Methuen.
D'AGOSTINo,R. B. (1971).An omnibustestofnormalityformoderateand largesize samples.Biometrika, 58, 341-348.
HUBBERT, M. K. (1967).Degreeofadvancement ofpetroleumexplorationin theU.S. Amer.Assoc.PetroleumGeologists
Bull.,51, 2207-2227.
KAUFMAN, of oil and gas fields.SPE/AIME Symposiumon
G. M. (1965). Statisticalanalysisof the size distribution
PetroleumEcon. and Evaluation,Dallas Section,pp. 109-124.
KAUFMAN, G. M., BALCER, Y. and KRYT, D. (1975). A probabilisticmodel of oil and gas discovery.In Methodsof
EstimatingtheVolumeofOil and Gas Resources.(J.D. Haun, ed.). Tulsa (Oklahoma): AmericanAssoc. Petroleum
Geologists.(Studiesin Geology No. 1.)
MOORF, C. L. (1962). MethodforEvaluatingU.S. CrudeOil Resourcesand ProjectingDomesticCrudeOil Availability.
Washington:DepartmentInterior,Officeof Oil and Gas.
ODELL, P. R. and ROSING, K. E. (1974).The NorthSea oil province:a simulationmodelofdevelopment.EnergyPolicy,
3, 316-329.
PELTO,C. R. (1973). Forecastingultimateoil recovery.Soc. PetroleumEng. J.,Paper No. 4261.
PHILLIPS, G. D. A. and HARVEY, A. C. (1974).A simpletestforserialcorrelation in regressionanalysis.J. Amer.Statist.
Ass.,69, 935-939.
RAIFFA, H and SCHAIFER, R. (1961). AppliedStatisticalDecision Theory.Boston: Harvard University.
RICHARDS, R. C. (1973). A three-dimensional model forpredictingnaturalgas reservesand availability.J. Canadian
PetroleumTechn.,January-March, pp. 41-46.
RYAN,J.T. (1973).An analysisofcrudeoil discoveryratein Alberta.Bull.Canadian PetroleumGeology,21, 219-235.
SHAPIRO,S. S. and WILK, M. B. (1965). An analysisofvariancetestsfornormality (completesamples).Biometrika, 52,
591-611.
VANEEK,W. H. (1974).Forecastingfuturehydrocarbondiscoveriesbased on a relationshipbetweenpast production
and reserves.Soc. PetroleumEng. J., PreprintPaper No. 4828.
WALKER, S. H. and DUNCAN, D. B. (1967).Estimationoftheprobabilityofan eventas a function ofseveralindependent
variables.Biometrika, 54, 167-179.
ZAPP,A. D. (1962). FuturePetroleumProducingCapacityof the U.S. Washington:U.S. Geol. SurveyBull., 1142-H.

DISCUSSION OF THE PAPER BY MR MEISNER AND MR DEMIRMEN

ProfessorA. G. HAWKES (University ofSwansea):In commenting on thispaperI findmyself operating


undertwo handicapswhichwillbe commonto manypeople presenttonight.First,I receivedthepapera
fewweeksago just as examinationpapers werebeginningto flowin, and whenI had also promisedto
completea longand rathercomplicatedPh.D. thesisexamination.Secondly,I have neveractuallyfound
an oil well in my life.Consequently,I shall restrictmyselfto makingsome superficialcommentsand
suggestions,withouthavinghad the opportunity to workany of themout in detail.
The art of statisticalmodellinglies in obtaininga balance betweena realisticrepresentation of the
physical phenomena being studied and a model which is both conceptuallyand computationally
sufficientlysimpleand robustto use as a workingtool. In forecastingproblemsespeciallythereis a
statisticalcreamingeffect, such that it is probablynecessaryto work extremelyhard to make any
significantimprovement on a good basic model.Theauthorsoftonight'spaperhavepresentedan elegantly
simplemodelwhichhas been thoroughly checkedin at leastone province,and I assume in otherswhich
have not been reportedhere.More complexmodelsprobablywould not do muchbetter and probably
would do worse.
The resultconcerningthe log normaldistribution (equation (2.2)), in whichthe actual log normal
distribution offieldsizesleads to a log normaldistributionofthefieldsizesactuallyfound,was rathernice.
However,I was ratherpuzzledbytheassumption(2.4)concerning theAparameter.Thisseemsto indicatea
changeas explorationcontinuesofthemannerin whichtheprobabilityofsuccessdependsuponfieldsize.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] Discussionof thePaper by Mr Meisnerand Mr Demirmen 23
It seemedto me more plausibleto fixA and to postulatethat,, the mean actual log fieldsize, should
decreasewithn insteadofremaining constant.In practicethismakesno difference becauseAcould thenbe
incorporatedintothe,Blratherthanintothe/2 in themodelspecifiedin equations(2.5) and (2.6).I do not
thinkthatis a practicalobjection,but I could not quite see the theoretical
justificationforit.
It appearsthatthemeanlogfieldsize,calledon inthispaper,and thelog-oddsratiofortheprobabilityof
successdecreaseina fairly
constantslope.Nevertheless, theremustbe a problemindecidingwhereto start.
My inclinationin modellingwouldbe to makethemodelsomewhatmorestochastic.For example,in the
model,ifwe have

in = log (1-O)
On
whereO,n
is theprobabilityofsuccessat thenthdrilling,
thenthepresentmodelcan be written
in theform
ln -ln-1 = (X2,

wherea2 is a constant.I would thinkof adding a littleextra,perhapssomethinglike the following:


ln 1n-1 = (X2 (X3Xn-1

whereO3 wouldprobablybe betweenzero and O2. Rememberthatxn= 1 ifthedrillingis successfulat the


nthattempt,so thatinsteadofhavingtheconstantslope thisis slightly modifiedaccordingto whetheror
not thedrillingwas successfulon thepreviousattempt.This allows thetrendto wandera little,makingit
slightlymore sensitiveto recentresults,and also making the original resultsfurtherback of less
importance.
Similarly,forthefieldsizes,at themomentthereis /n, whichis themean ofthelogarithmand thus
appears in the log normaldistribution.
The presentmodel can be writtenessentiallyas follows:
1lt)n 1)n-1 = f29
where/2 is a constant.Rathervaguely,I postulateperhapswe mightwrite+ (some suitablefunctionof
recentfieldsizes).That would have the same kind ofeffect on thisprocessas theextratermhas on the
probabilityof success(shownabove).
In particular,
one ofthefeaturesofthemodelis theseparationoflookingat theprobability ofsuccessat
all and lookingat thefieldsize obtainedifsuccessis achieved.Thishelpsto makethemodelsimple,butitis
one ofthefewassumptionsthatdoes notseemto havebeencheckedwiththedata-as faras I could see,at
any rate,in thepaper,althoughit mayhave been done somewhere.I wonderwhetherthisis right,and
whetherperhapsiftherewas a runofno successesthatthisdid notprovidesomeinformation about thesize
of the potentialfieldsas well as whethertheywerezero or non-zero.
It maybe thatthereis somekindofcorrelationbetweenthesetwoprocesses,theprobabilityofsuccess
and the fieldsize question.Perhaps a link,if thereis one, mightbe establishedthroughthissuitable
function, whateverit mightbe, added to theexpressionfor )n-n-l1
The MonteCarlo procedurein Section4 is a nicesensiblewayofhandlingthepredictivedistribution,
whichwould be verydifficult to cope withanalytically.However,I wonderhow largeN, thenumberof
trials,so-to-speak,has to be to providereasonablypreciseinformation about the presumablyrather
importantuppertail.The smoothcurvein Fig. 11 musthave been obtainedusinga verylargeN, or a
smoothingtechniquesuchas a kerneldensityorhistosplinemethod.I wonderwhether itwouldbe usefulto
tryto fitsome reasonableclass ofdistributions to describethesepredictivedistributionsof recoverable
reserves.The parametersofthisclass wouldmosteasilybe estimatedfromsimulateddata. The idea ofthis
is thatwitha fittedcurveitmightbe possibleto be rathermoredefinite about theuppertailbehaviourthan
is possiblewithdirectsimulation.
I was notclearabout thelinkbetweentheestimationproceduredescribedin detailin thepaper,based
on past results,and the information available throughgeophysicalsurveysand so on about individual
prospects.I assume thattheformeris used forstrategicforwardplanning,withthelatterused formore
immediatedecisionson whereto drillnext-although I gatherfromtonight'sverbalpresentationthat
perhapsitis simplyusedas somekindofa checkon geologists'estimation.Butthereis no directlinkingof
the two things.
I cannothelpfeelingthatthepresentation ofthispaperis a littlelikewatchingthedance oftheseven
veils.The authorshave takenoffone veil,and thathas provedquiterevealing.In Section1.3,wherethey
talkaboutgeophysicalevidenceand evaluatingindividualprospectsand so on,I thinkthattheyhavelifted

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
24 Discussion of the Paper by Mr Meisner and Mr Demirmen [Part 1,
thecornerofa secondveiland hintedat evenmoreinteresting I can see decision
possibilitiesunderneath.
theoristslickingtheirlips at theprospect.In viewofthefinancialrewardsinvolved,I doubt ifall willbe
revealedjust yet.Nevertheless, theauthorsare to be congratulatedon presentinga usefuland interesting
paper on a matterof obvious importance.I have greatpleasurein proposingthe vote of thanks.
Mr A. W. DAVIES (BritishPetroleumCo. Ltd,London): ProfessorHawkes has dealtwiththe"number
crunching"aspectsof thepaper.I willgivea fewcommentson thegeologicalside. If two geologistsare
asked to make a report,therewill be the reportand two minorityreports,so perhapstherefore it is
fortunatethat Dr Demirmenis not heretonight.
My mainconcernis about thedefinition ofthe"petroleumprovince",particularly wherethereis the
statement(in Section 1.1) thatthisdoes not necessarilyreferto a geologicalprovince.Bearingin mind
unemployment in Britainat the moment,and beinga geologist,I findthatslightlyworrying.
I accept that in the examplecited the methodworks.I thinkthatit worksbecause the basin, the
geologicalprovince,chosenfortheexample I do not know how manyotherexamplesweretried-is a
homogeneouslyheterogeneous basin.Wherethegeologywillbe morecomplex,whichis themoreusual,I
thinkthatthemethodcould notworkfora geologicalprovinceas such,butwould have to be limitedto a
geological horizon withina geological province.I'll tryand exemplifythat withrespectto the U.K
ContinentalShelf in theSouthernNorthSea in thethegas playarea oftheUK and ofHolland virtually
all theplayshave thesame sourcerock,thesame reservoirrock,thesame seals,thesame structure, age of
structuralgrowthand timeofgas migration.All theseare similar,in such a situation,thissystemwould
work.Contrastwiththat,the NorthernNorthSea and the CentralNorth Sea whereit is certainlynot
homogeneouslyheterogeneous, wherethereare at least six sources,15 differentreservoirsand different
horizons and, ifthereare 15 different reservoirstherewill have to be 15 seals. There are all typesof
structures quitea lotmorethanthethreeexamplified [bytheauthors].Agesofgrowth, migration and re-
migration:all thesechangethesituationsuchthatitis notpossibleto deal withtheseareas as a geological
province.It wouldbe necessaryto take,say,thechalkhorizonsofNorwayseparatelywhenthismethod
may well work but it is not possible to throwin the chalk discoveriesof Norway withthe Jurassic
discoveriesof Norway it is like apples and oranges.
The industry drilledquitea numberofwells 33 in Norwaybeforefinding anything. Therewas then
successand,as is thewont,whenwe havesomethingwechaseit,so a similarsortofplaywas chased.It was
quitesome timelaterthatthedecisionwas takento drilldeeper whichled to findingthenewsequence.
Thus,therewere,in effect, two learningcurves sometimeswe are ratherslow in appreciatingthatthere
is more than one play.
This methodmayworkundercertaincircumstances, suchas in theexamplecitedwhereitdoes appear
to work.But thoseconditionsin whichit willworkare thoughtto be rare.It is veryrarenow foranyone
companyto have all therawdata availablein,say,a geologicalprovince,whetheritbe thewholethingor
just one horizonin that province.Perhaps in Nigeria Shell will have all those data, but normallyoil
companiesare unable to get all the data in a province.Stickingby my argumentthatit has to be an
homogeneouslyheterogeneous basinforthemethodto work,we cannotaffordto have a lot ofdata from
onlyone partofthatbasin.Ifall thesampleor a sufficiently largeaerialsampleis notavailable,theanswer
may be anomalous.
Thus,thefirst thingis to havethedata. Secondly,therehas to be availabletherawdata,so thatthedata
can be interpreted consistently and on a compatiblebasis. Equally, it is difficult
to findan area where
politicshave not affectedthe oil industry, eitherin timeor in access. Then thereare the difficulties of
definition. Whatis a discovery? Is itcommerciality?If50 millionbarrels[reserves]werefoundin say 1973
in theNorth Sea we would have thoughtthatwas a good find,but we would have gone on to look for
somethingbigger.Now, however,we are lookingfor50 millionbarrelsof oil reserves.How is thatwell
drilledin 1973viewed?Is thattheactualdiscovery, or is thediscoverythewellthathas beendrilledin 1980
to make sure thatthe 1973 discoverywas right?
Turningto thesize ofthediscovery, thisindeedis difficult.It is commonlyunderestimatedalthough
certainlynot always,overestimation may take place forstock pushingpurposesamong certainother
countries.Or perhapsforpolitics.Thereis a case in thisregardin whicha certaincompanywas havinga
fight, or an acrimoniousdebate,withone government. It thenfounda largediscoveryin anothercountry
and,to putpressureon thefirstgovernment, announceda verylargediscoveryin thatothercountry.It is
not possiblealwaysto guaranteethefiguresthathave been published.Of course,thereare also genuine
mistakes.There are cases of deliberateunderestimation, not just forconservativereasons but forthe
followingreason.In theStatesa bonuscan be gainedfora discoveryon a newstructure, so ifsomeonecan

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
19811 DiscussionofthePaper by Mr Meisnerand Mr Demirmen 25
persuadetheauthoritiesthaton thesame mainfeaturethreeseparateoil or gas accumulationshave been
discoveredhe can receivethreediscoverybonuses.
Otherfactorsthataltertherateofdrillingincludethefollowing: a companymightgo intoan area and
notbe verysuccessfulfora fewyears,so it markstimeand hopes thatsomeoneelse willbe successfuland
once someoneelse is successful, thefirstcompanyfollowstheleader.Once a companyis successful, starts
productionand cash beginsto flow,and it getsthetax dollar or thetax pound now it is muchmore
preparedto drillon a geologicalhunch,a gutfeelingor whateverifit is now spending30 pence-?srather
than ?1-?s.
In anybasin,unlessitis theveryrareoccasion whenonlyone companyis workingand has thewhole
basin to deal withitself,I findit verydifficult
to see how thismethodcan work.That is speakingas a
geologist,realizinghow dubious are some of the data givento our statisticiansby the geologists.
One oftheinteresting thingsofthismethodis thatthelearningcurveis eliminatedfromthestatistics
but when21 yearsdrillingin a particularcountrywithno successis eliminatedit is a verylong learning
curveand a greatdeal ofdata lost.For instance,Norwayhad 33 wellsbeforethegeologistsdecidedthat
anotherplay had betterbe tried,whereason the North Slope of Alaska the thirdwell got 10 billion
barrels and thatis on thelearningcurve,yetitis muchmorematurethanmostbasinsintheworldexcept
forthe Middle East and the Gulf.
My pointreallyis thatthedata givento thestatisticians to workon by any practicaloil man are of
dubious quality sometimeshighlydubious,sometimesslightlyless so. It is importantalwaysto bear in
mindthat in goingto the greatand accurateextremesthat one appears to withnumbersthe original
numbersare dubious.I am sorrythatI am usingtheword"dubious"ratheroften,butI thinkthatitis the
truth.
The comparisonmade by theauthorsbetweena lotteryand theexplorationbusinessis correct.It is a
usefulanalogy,exceptthatin lotteriessomeonealwayswinssomething; in theoil businessitis possibleto
spend $100 millionand get nothing.
In summary,this mathematicalapproach has its value to individualcompanies in certainrare
circumstances wheretheyhave all thedata,whichhave all beeninterpreted bythecompanies'ownpeople
andcareconsistentand wheretheyhavemadesurethatthedata deal onlywitha geologicalprovincewhich
is rathersimple,or withone geologicalhorizonwithinone province.
Mathematically, I haveno argumentwiththepaperbutI thinkweshouldbe awareofthedata givenon
whichto work.
Finally,I wouldliketo second ProfessorHawkes' voteofthanksto theauthors,Mr Meisnerand Mr
Demirmen,and particularly to Mr Meisnerforpresentingthe paper.

The vote of thankswas passed by acclamation.

Mr G. J.A. STERN(I.C.L.,London):I cannotcontribute muchto a theoreticalevaluationoftheauthors'


method,butI wouldliketo be able to assessitsuse as an aid to management, and to do thisI askjust which
resultsfromthe methodmanagersuse, and how theyuse those results.I ask also how accurate the
confidenceintervalscalculatedhaveprovedinpractice,as I feelthatwhatmanagersoftenwantis whathas
been calculatedhere:a measureofexpectationand a confidenceinterval.In mycommentsI make some
assumptionsas to whatmanagerswant,and suggestthatcrudermethodsmaygivethemwhattheywantin
a moreconvenientform.Firstly, theauthorsestimatec,thechanceofmakinga further discovery, whichfor
thelast40 drillingsis calculatedas 0 993. I suspectthatthemanagerjustwantsto knowwhetherthereis a
nearcertainty ofanotherdiscovery, or a nearcertainty ofno discovery,
and all valuesofc between0 05 and
0 95 are thesame to him.Ifthisis so, I would say thatit seemsthatthegraphofdiscoveriesis tendingto
about 0 2 per drilling,so that the chance of one or more in 40 drillingsis somethingof the orderof
1 - exp(-8) = 0 9997, and thatfromthe manager'spointof view thisanswerdoes not differ fromthe
authors'0 993,but is calculatedby a methodhe can believethathe understands, and can manipulatefor
further privatecalculationsof his own-calculations whose naturehe may not wishto discuss witha
statistician.
Again,a graphbased on Table 1 suggeststhatrecoveriespertendrillingsare levellingout at
about 75 millionBBLS, or 300 for40 drillings, close bothto theforecastneartheend ofSection4, and to
the observedoutturn.
As theauthorsthemselvessay in Section5, and as theSecondersuggested,the"observed"recoveries
are by no means accuratelyobserved,and such a graphicalmethodat least lets the managermake
allowancesforsuch inaccuracy,and giveshim a view of the possible varianceor confidenceinterval.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
26 Discussionof thePaper by Mr Meisnerand Mr Demirmen [Part 1,
To estimatetheconfidenceintervalmoreclosely,I regressedrecoverieson a second-orderpolynomial
in wellnumber(groupedin tens),usingforthisPackage X, a conversationalstatisticalsystemwhichthe
CentralStatisticalOfficewerekindenoughto letme use.This gave a residualstandarderrorforgroupsof
tenwellsof325.Approximately, thiswouldsuggestan R.S.E. of650 forfourobservations, 40
representing
wells,and assumingnormality(whichis by no means totallyjustified)I would take somethinglike
300 + 700 forthe67 percentconfidenceinterval.I wouldalterthisto 0 to 1300to allow fornon-normality
caused partlybytheimpossiblenatureofvaluesbelowzero.I acceptthatsomethinga littlemorescientific
thanthisis needed,butI suggestthatthefinalanswerwouldstillbe close to this:expectationaround300,
67 percentintervalof0 to somewherebetween1000and 2000. I wouldsuggestthatsomesuchmethodas
this can be more usefulin many contexts,because it is easier to see how one can make at least an
approximateallowance forknowninaccuraciesin the data, and because it is easier to manipulatethe
calculationsfurther to make otherestimates,e.g. discountedvalues offuturediscoverieswithconfidence
intervals.It mayoftenbe truetoo thatthemanagerwillhavemoreconfidence in suchcrudemethodswhich
he partlyunderstands thanhe willhavein methodswhichhe cannotsparethetimeto getto know.I would
concludehoweverbyemphasizingthatI don'tknowwhethermypictureofwhatmanagerswantis validin
thiscase, and look to theauthorsforenlightenment, and also I wouldexpressthehope thattheirmethod
has been or willbe presentedto such usersin a formwhichmakes it seem moreaccessibleand easy to
adjust. The foregoingis myown opinion and does not involvemycompany.

Dr F. H. HANSFORD-MILLER (InnerLondon EducationAuthority): I also wouldliketo raisethesubject


of the petroleumprovinces.Sub-section1.1 of the Introductionof the paper reads as follows:
"For practicalpurposes,nationalfrontiers can be consideredto be provinceboundaries."
What,then,is the positionon the North Sea continentalshelfwithregardto the national boundary
betweentheUnitedKingdomand Norwegianoil zones?This frontier line,at present,is drawn,according
to theUnitedKingdom-NorwegianTreatyof1965,on themedianprinciple.It so happens,therefore, that
thelinestraddlestheStatfjordand Friggfields,so thatinthiscase thenationalfrontier cutsrightacrossthe
geologicalprovinces,about whichwe have heardin thediscussion.The 1965 Treatycompletelyignored
theNorwegianTrough,ofdeep waterup to 500 metresinplacesand thusnotcontinental and I am on
shelf,
record(Hansford-Miller, 1980) as havingpointedout that,in my view,the United Kingdom oil zone
shouldextendeastwardsto theedgeofthecontinentalshelf,i.e.to theNorwegianTrough,and ifthiswere
to happenthewholeofStatfjord and Friggwouldbe in UnitedKingdomwaters.How wouldthisaffect the
findingsoftheauthors?Withregardto thegiantStatfjordfielditis estimatedthat11-15 percentis in the
United KingdomZone, withthe restin the NorwegianZone. If provincesin that area are taken into
account surelythatmustaffectthe conclusionsof the paper.
Secondly,we have recentlyhad a declinein theestimatesoffutureyieldsfromthe UnitedKingdom
NorthSea zone. Mightthisbe due to theauthors'forecasts, or is it possiblydue to thecreamingprocess
discussedin the paper?

Mr C. J. TAYLOR (O.R. Department,BritishGas Corporation):First,I would like to add my


congratulations to theauthorsforan interestingpaper.WhenI readthepaper,whatworriedmemostwas
thisquestionof the samplingaspect of drilling.In Section2 it states:
"Ifa discoveryis madein randomdrilling, thatthefieldsizewillliebetweenvand
theprobability
v+ dv will be proportionalto v '"
It was thereference paperthatworriedme because drillingis generally
to "randomdrilling"in thewritten
speakingcarriedouthavingsomeinformation fromgeologicalmaps and seismicsurveys.Whatwouldthe
effectof thisbe on the results?The problemhas been recognizedby the introductionof a i value that
changeswithfieldsize accordingto an empiricalrule,and ifthisadequatelycoversthemathematics ofthe
situationthenI am satisfiedabout it.
However,secondly,thereis thequestionthatifone wellis drilledand thenanotherone is drilledclose
by,the second drillingmighthit the same field(as the first).Under randomdrillingor otherwisethis
presumably is somethingthathappens.I do notknowwhatwouldbe theeffect ofthaton themathematics.
Thirdly,about theauthors'experienceon thegas sideofthesubject,could Mr Meisnertellmewhether
he has analysedany data, in whichboth gas and oil are given,as measuresoffieldsize? If so, would he
recommendtreatingthetwo separately,or would he tryto predictsome sortofbivariatedistribution in
forecasting?

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] Discussionof thePaper by Mr Meisnerand Mr Demirmen 27
Dr S. J. TAYLOR (Lancaster University):Tonight's proposer has suggestedthat the quantities
n= log{(10 - OJA)/ should have more structurethan the authors' model, mn - q -1 = X2 There is
evidencein the paper favouringhis suggestion.A cursoryinspectionof the 22 pointsplottedon Fig. 5
showsfirstly thatno consecutivepairsofthe"probability ofsuccess"valuesareequal and secondlythatthe
sequence of thesevalues has negativelyautocorrelatedfirst-differences.
These twoobservationscastdoubton theauthors'modelfortheq,,althoughI havenotyetperformed
significance themeeting,
tests.(After I calculatedthefirst
autocorrelation coefficientto be - 060, usingthe
This coefficient
21 first-differences. at the 1 per cent level.)
is significant

Dr A. C. ATKINSON (ImperialCollege,London): I would liketo congratulatetheauthorson a paper


whichprovidesan interesting exampleofstatisticalmodellingthatavoids overcomplexity butappearsto
produce usefulanswers.
I have two small commentson thedata. In Fig. 5 thereappears to be a see-saweffectindicativeof
negativecorrelation.On theassumptionofconstant0 a 2 x 2 tableofpairsofsuccessiveresultsyieldsa %2
value of 0 73. There is no reason to believe that an analysis using the logisticformfordecreasing
probabilitieswould giveappreciablystrongerevidenceofnon-independence. Thus theauthors'untested
assumptionof independencein (3.1) appears vindicated.
The firsttwocolumnsoffieldsizes givenin Table 1 appear plausible,but I am not so sureabout the
thirdcolumn.Severalofthesizes seem rounded,such as 20, 100, 125. In contrastto thesethereare five
appearancesof 8 8. Do the authorshave any commentson thesenumbers?

ProfessorD. R. Cox (ImperialCollege,London): I enjoyedthispaper verymuch.The temptationto


suggestcomplicationsto authors' models has to be resisted.Nevertheless,the exponentialultimate
pessimism(On-+0, Vn-+ 0 as n -+ oo) ofthemodelsmightforlong-term be a disadvantageand
forecasting
extraparametersmightbe included.Of course,the quantitiesin a given"province"are finite,so that
ultimatelythe pessimismis right,but the rate of approach assumedhereis ratherstrong.
-TheanalyticalcalculationofthepredictivedistributionofRnm,whetherby Bayesianor non-Bayesian
methods,is a challengingproblem.I do notdoubtthattheauthorswerecorrectto worknumerically butI
suspectthatsomeprogresscould be made analytically expansionofthedistribution
eithervia asymptotic
forfixedparameters, or byjudicious fittingby moments,followedby perturbationto allow forerrorsof
estimationof the parameters.

werereceivedin writing,
The followingcontributions afterthe meeting.

Mr L. A. BAXTER(CentralElectricity GeneratingBoard): The authorsmake use of a Monte Carlo


simulationprocedureto obtain values of the distribution
of
m
Rnm =Y rn+j,

wherern+j (j = 1,2,... m)followsa lognormaldistribution withprobability0n+j a mass of probability


1- On+j beingassignedto the origin(Section4). I would like to draw the authors'attentionto a more
methodofcomputingthedistribution
efficient ofsumsofindependently non-negative
distributed random
variables.
Clerouxand McConalogue(1976)presentan algorithm forgeneratinga cubicsplineapproximationto
any probabilitydistributionfunctionF E C2[0, oo) (see also McConalogue, 1978).Briefly,theabscissa is
dividedintoa seriesofpanelsofequal width,h say,and F is approximatedby a cubicin each panel.This
representation,togetherwithits firstand second derivatives, is continuousat each node, and hencean
extremelysmooth approximationis obtained. An arbitrarilyclose approximationcan, withinthe
limitationsimposedbyround-off error,be achievedbychoosingh sufficiently small.The derivativeofthe
cubic splineapproximatesF', and henceevaluatingthe Stieltjesconvolutionofdistribution F
functions
and G,viz.

F * G(t)= F(t-u) dG(u),

is reducedto thestraighforward
taskofintegrating The algorithmis particularly
polynomialsnumerically.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
28 Discussionof thePaper by Mr Meisnerand Mr Demirmen [Part 1,
convenientforcomputingrecursively-defined
convolutionsof the formF(n)(t),where
Ffn() F(t), n=I
(n- 1) * F(t), n> 2.
The accuracyof the approximationdoes not diminishappreciablywithincreasingn or t.
I have useda generalization
ofthisalgorithm(McConalogue, 1980;see also Baxter,1980)numerically
to convolutea varietyofprobabilitydistributions,
in particularthelognormal,Weibull,gamma,inverse
Gaussian and truncatednormaldistributions.I havefoundthatthismethodis considerablymoreaccurate
than Monte Carlo simulationand is also moreeconomicalon computertime.

Mr E. A. FIELD (Esso PetroleumCo Ltd): The proposed procedurefor forecastingoil and gas
discoveriesrestson assumptionsmade about (i) successrateand (ii) fieldsizes.Obviouslythesuccessrate
willdepend upon two factors(a) explorationefficiency and (b) area potential,and some attemptmustbe
made to determine these.The purposeofthenewprocedureis to contributetowardstheimprovement in
explorationefficiency,and thiscan onlybe determined bycomparingthelevelofefficiency beforeand after
theintroduction ofthenewprocedure.Cozzolino addressedthisproblem(Cozzolino, 1979)in hispaperto
the AIME. In predictingfuturesuccess the dilemmais to resolvethe difference betweena highpast
explorationefficiency in a low potentialsearcharea and a low past efficiency in a highpotentialsearch
area. The former implieslimitedfuturesuccesssincemostfieldshave probablybeenfoundalready,whilst
the latterimpliesthatmanyfieldsremainundiscovered.
Regardingfieldsize distribution thechoice of log normalhas manyprecedents.The famousFrench
studyoftheAlgerianSahara (Allais,1957),evenwentso faras to say"As faras wecanjudge thelog normal
distributionis one ofthefundamental laws ofnature".For any assumptionabout fieldsize distribution,
even random drillingwould demonstratethe phenomenacalled "creaming".Extensive"hindsight"
analysisof explorationstrategieshas been reported(Haraugh et al., 1977) includingrandom drilling
(Drew, 1974).
It is encouragingthatworkofthistypeis stillproceeding,forin an era ofenergycrisis,each unitof
information thatcan be wrungfromthe data carriesa veryhighreward.

The AUTHORS
repliedlater,in writing,
as follows.

We would firstliketo thankthediscussantsofour paperfortheirinterestand usefulcomments.The


vastmajorityofthesecommentsareconcernedwith,orare relatedto,thestructure ofourmodelwhichwas
mainlydetermined byitspurposeas a meansofforecasting and our beliefsabout theprocessunderlying
thecreamingphenomenon.As outlinedin theintroduction, our procedureis used forstrategicforward
planningand aimsat forecasting futurediscoveriesfroma numberofexplorationwellsyetto be drilledand
not to decideon whereto drillnext,as correctlyunderstoodby ProfessorHawkes. We agreecompletely
withMr Field thatthe successratedependson area potentialand explorationefficiency. However,we
believealso thatthediscoverysequenceoffieldsizes is an importantindicatorofexplorationefficiency,
sinceit is in theinterestoftheoil industryto findthelargerfieldsin theearlierstageofexploration.The
area potentialis characterizedof course by the numberof oil/gasfieldsin that area (discoveredand
undiscovered) and thedistributionofthefieldsizes,whichweassumedto be log normalwithas parameters
themean log fieldsize ,uand thevarianceofthelog fieldsize U2. We regretthatin our paperwe obviously
did not succeedin makingclear thesignificance ofthedeclinein Aas a measureofexplorationefficiency
since ProfessorHawkes was ratherpuzzled by our assumption(2.4) concerningA. Let us considera
hypotheticalpetroleumprovincewith only one type of oil reservoirwhose areal extentis directly
proportionalto itsfieldsize. In thecase ofrandomdrillingthevalue ofAwillbe equal to one since,given
success,theprobability ofobtaininga fieldsizebetweenvand v+ dvwillbe proportionalto itsarealextent
whichin its turnis directlyproportionalto v. In scientificdrillingthroughthe use of geology and
geophysics,we are expectedto do better.So we hope that in the earlystage of explorationA will be
considerablylargerthanone and willdeclineto valuesbelowone lateron. In a realpetroleumprovinceAis
unknownand varieswiththetypeofreservoir, but some averagevalue could be estimatedbya loglinear
regressionoftheareal extentofthefieldsize ifthesedata wereavailable. An indicatorof thenumberof
oil/gasficldsis thenumberofpromising structuresor trappingconfigurations thatmightcontainoil orgas,
of whichonlya fractionwill turnout to be a success.
Fromtheabove itmaybe clearthatwedo notexpect,as suggestedbyMr Field,thatourprocedurewill

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] Discussionof thePaper by Mr Meisnerand Mr Demirmen 29
contribute towardstheimprovement ofexplorationefficiency.On thecontrary, thebasis ofourprocedure
is projectingthe observedpast explorationefficiency into the future.
We hopealso thatwe havenowtakenawaytheworriesofMr C. J.Taylorabout randomdrilling. It will
be clearthatrandomdrillingis onlyconsideredas a reference case againstwhichtheefficiency ofscientific
drillingcan be judged. A successfulexplorationdrillingis usuallyfollowedby one or more appraisal
drillingsto assess thefieldsizeand plan thedevelopmentofthefield.Hence theprobabilityofhitting twice
thesamefieldis extremely lowand wouldbe consideredas a failurein ourprocedureifitreallydid happen.
In Section 1.1 the problemof forecasting oils and gas simultaneously is brieflydiscussed.
We greatlyappreciatethecommentsofProfessorHawkeson theartofstatisticalmodelling.Therefore,
inourturnwe area bitpuzzledbyhissuggestions to makethemodelmorecomplicated.Writingthemodel
forOn in the formhe suggestsyieldsfinally:

on= {I +exp Q + o2n+1 3 E X)}Xi

wherehe expectsa3 to be negativeand in absolutevaluebetweenzeroand ?C2 The modelforOninthisform


maybe usefulforone stepahead forecasts, butitis notveryhelpfulforforecasting thesumofmstepaheads
whenmis ratherlarge,sinceexceptforthelast termthewholesum to be forecastedis now a partofthe
second indicator.So, we are happy with ProfessorCox's commentthat the temptationto suggest
complicationsto our model has to be resisted.
Dr S. J.Taylorseemsto have a strongpointin hisremarksconcerningthenegatively auto-correlated
first-differencesinFig. 5,althoughthecommentsofMrAtkinsonseemto weakenhispointagain.Indeed,if
we assume a constantprobabilityof success,then the data do not yield any evidence against the
assumptionofindependence. However,wedo notagreewithMr Atkinson'sremarkthattheassumptionof
independencehas notbeentested.On p. 20 we gave a shortdescriptionofour analysisofresidualsand we
concludedthattheresultsofthisanalysiswerequiteconsistentwithwhatcould be expectedon thebasis of
theassumedbinarydata generatingprocess.Of course,one can argueabout whatthemostappropriate
testis in thiscase,becausewe are dealingwitha binarydata generating processwhoseparameteris a non-
linearfunctionof the indicator.Let us considerthe data in Fig. 5 as resultingfroma binomialdata
generatingprocesswithsample size equal to 10 and wherethe probabilityof successfollowsa logistic
declinewiththe midpointof the numberof explorationwells drilled.Let us further assume that the
binomialdata generating processcan be adequatelyapproximatedby a normaldata generating process
withthesame meanand varianceand thatan adequate approximationofthepredictivedistribution ofa
one step ahead forecastcan be obtained by means ofa order
first linearapproximation. Then we can use
heretoo themethodoftherecursive residualsas describedin ourpaperfortheloglineardeclineinfieldsize,
whichyieldsa vonNeumann'sratioequal to 2 65,whichis notsignificant at thetwo-sided10percentlevel.
Hence we do not thinkthat Dr S. J. Taylor's firstautocorrelationcoefficient constitutesthe most
appropriatetestsinceno allowanceis made fortheheteroscedasticity oftheresidualsand thecorrelation
betweenthemintroducedby the use of the commonestimate&.
The commentsof Mr Field on area potentialand explorationefficiency answermore or less the
questionofProfessorHawkeson thecorrelationbetweensuccessrateand fieldsize givensuccess.Finally,
withrespectto themodelstructure, ProfessorCox is quite righton theultimatepessimismbuiltintoour
model,but in our beliefthisultimatepessimismis in agreementwiththe generallyobservedcreaming
phenomenon.For example,accordingto our modelthetotalprobabilityofa discoverywitha fieldsize of
at least50 106 BBLS whendrillingthe500th explorationwellinprovinceXX 11,is roughlyequal to 10 -4. In
our beliefpessimismofthisorderofmagnitudeis justifiedgiventheexplorationhistoryofthisprovince
unlessadditionalinformation becomes available.
The commentsof Mr Davies and Dr Hansford-Miller concernour definition ofpetroleumprovince.
Mr Davies is absolutelyrightin pointingat thedifficulty ofdefininga "petroleumprovince".The more
restricted ina geologicalsensethemorehomogeneousand hencebetterchancesthatthecreamingprocess
willbe observed,and can be usedfruitfully forprediction. We adherein practiceas muchas possibleto the
"playconcept"wherethesame sourcerockfeedsthesame reservoirformation capped by thesame seal.
However,whenthisprovesimpracticable we mayhaveto dilutetheserestrictions, usuallyat a price.When
mixinggeologicalapples and orangeswe normallygeta warningfromtheanalysisprogrammethatno
significant creamingis observed.In suchcases it becomesobviouslyimpossibleto make a forecast.This
formsas ifit werea built-insafetydeviceavoidingunrealisticuse ofthemodel.In intermediate cases the
degreeofdeclinein successrateand/orfieldsize maybe so weak thattheforecastalso willbecome very
vague (in termsof its predictivedistribution) but may stillbe of some value.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
30 Discussionof thePaper by Mr Meisnerand Mr Demirmen [Part 1,
Under-estimation offieldsizesaredealtwithoutsidethemodel.Field size"appreciation"curvescan be
generatedin provincesand usedto correctrecently declaredreservesina statisticalway,thusreducingthe
possiblebias in the forecasts.
We agree withMr Davies' conclusionsthatour procedureis dependenton the availabilityof well
organizeddata bases and therefore notofimmediateuse to everybody, althoughthetrendtowardspublic
data bases is clear.For example,thePDS systemin theU.S.A. could providemuchofthedata requiredfor
thiskind of analysis.
Withrespectto Dr Hansford-Miller's comments, thestatement in Section1.1ofourpaperis notmeant
to implythat national frontiers are our preferred bounds to a petroleumprovince.However,when
analysingthewholecreamingprocess,thenationalfrontiers havea significance becausetheymayseparate
areas wheretheoil industry had greatfreedomto explorefromotherswhereexplorationhas beenseverely
slowed down by the Government.So, althoughthenationalfrontiers splita geologicalprovincerather
arbitrarily, it avoids mixingapples and orangesin anothersense.
Declinesinestimatesoffuture yieldsovertheyearscan resultfromthe"acountingeffect", i.e.transferof
discoveredoil or gas fromestimatedfuture discoveriesto provenreserves. Anotherfairlycommoncause is
thatpublishedestimatesdo not necessarilyrepresent mean valuesin thesenseofstatisticalexpectations,
butsomeother"optimistic" percentile distribution.
ofa predictive Withincreasinginformation thesekinds
of estimatesbecome adjustedmore oftendownwardthan upward.
Mr Sternasks whichresultsfromour procedureare used by management, how theyuse theseresults
and whethermanagershave confidencein crude"on the back of an envelope"methodsratherthan in
sophisticatedprocedureslikeours.In Shelltheusual wayto presentestimatesofuncertainquantitiessuch
as oil and gas reserves,is bymeansofexpectationcurvesas discussedin Section1.3ofourpaper.Managers
can use thesecurvesto read fromthemtheyard-sticks whichare relevantto a givendecisionproblem.
Dependingon thecontextthismaybe a givenpercentile, themean or themode.In industrialpracticethe
resultsof studiesby geological/statistical experts(posteriordistributionsof parameters,expectation
curves,etc.)are givento and discussedwithplanners(how likelyis a possiblescenario?)and economists
(how much value can one attach to a certainarea?) In this way an interfaceis formedbetweenthe
geological/statistical expertsand the decisionmakers.
For thiskindofproblem"managers"do nothave thedata northetimeto makeevencrudestatistical
estimates.Managementemploysexpensivespecialiststo do thejob of summarizingall of theavailable
information in termsof an expectationcurveshowingall the uncertainties. Such an expectationcurve
automaticallytellsthemanagerhow fuzzythefutureis (see Fig. 2). In our procedurethenoiseinherent in
the creamingprocessis faroutweighingthe inaccuraciesin the data.
Finally,therewerea fewquestionsand commentson computationalprocedures.We are grateful to Mr
Baxterfordrawingourattentionto a moreaccurateand efficient methodofcomputingthedistribution of
sumsofindependently distributed non-negative randomvariables.We certainly willstudythepotentialof
thismethodforourproblem.However,theproblemis morecomplicatedthanas presentedin Mr Baxter's
comments.The problemis notto evaluatenumerically, eq. (4.1)in our paperbutthenumericalevaluation
ofthepredictive distributionofRn,m along thesteps(1) and (2) as describedin Section4 at thebottomofp.
16.For example,one complicationis thattheestimatesofthern+j are notmutuallyindependent sincethey
are all based on a commondata set.
We would welcomeany progressin the directionof an analyticalsolutionof the evaluationof the
predictive distribution ofRn,m. For thetimebeing,however,we viewa MonteCarlo simulationprocedure
followedby a simplesmoothingof the resultingcumulativedistribution functionas an effectivealbeit
perhapsnot the mostefficient solution.

REFERENCES IN THE DISCUSSION


Algerian
ALLAIS, M. (1957). Method of appraisingeconomicprospectsof miningexplorationover large territories:
Sahara case study.Manag. Sci., 3, 285-347.
BAXTER,L. A. (1980). Some remarkson numericalconvolution.Commun.in Statist.B, (to appear).
Paper deliveredto the 8th
COZZOLINO, J. M. (1979). Measurementand projectionof explorationsearchefficiency.
HydrocarbonEconomicsand Evaluation Symposium,A.I.M.E., February1979.
convolutionintegrals
CLEROUX, R. and MCCONALOGUE, D. J.(1976). A numericalalgorithmforrecursively-defined
involvingdistribution functions.Manag. Sci., 22, 1138-1146.
of resourcebase exhaustionvia a
DREW,L. J. (1974). Estimationof petroleumexplorationsuccess and the effects
simulationmodel. US Geological SurveyBulletinNo. 1328.

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions
1981] Discussionof thePaper by Mr Meisnerand Mr Demirmen 31
HANSFORD-MILLER, F. H. (1980). Territorialclaims in the North Sea. The Times,12 May 1980; also: North Sea
boundarytreaty.The Times,16 June1980.
HARBAUGI J.W., DOVERTON, J.H. and DAVIES, J.C. (1977). ProbabilityMethodsin Oil Exploration.London and
New York: Wiley.
MCCONALOGUF, D. J. (1978). Algorithm102. Convolution integralsinvolvingprobabilitydistributionfunctions.
ComputerJ.,21, 270-272.
(1980).Numericaltreatment ofconvolutionintegralsinvolvingdistributions
withdensitieshavingsingularities
at
the origin.Commun.in Statist.B, (to appear).

This content downloaded from 200.130.19.157 on Wed, 13 May 2015 15:49:34 UTC
All use subject to JSTOR Terms and Conditions

You might also like