S

SPE 1613
300

P
Probabilis
stic Decline Curve
e Analysis Reliably Quantiifies Uncertainty iin
S
Shale Gas Reserv ves Regarrdless of Stage off Depletio
on
R
Raul Gonzalez, SPE, Xingllai Gong, SPE
E, Duane McV
Vay, SPE, Te
exas A&M University

C
Copyright 2012, Society
y of Petroleum Enginee
ers

T
This paper was prepare
ed for presentation at the SPE Eastern Regio
onal Meeting held in Le
exington, Kentucky, US
SA, 3–5 October 2012..

T
This paper was selected for presentation by an a SPE program comm mittee following review of information containned in an abstract submmitted by the author(s)). Contents of the papeer have not been
re
eviewed by the Society y of Petroleum Engine eers and are subject to o correction by the autthor(s). The material ddoes not necessarily reeflect any position of the Society of Petroleu
um Engineers, its
officers, or members. Electronic
E reproductionn, distribution, or stora
age of any part of thiss paper without the wrritten consent of the S Society of Petroleum E Engineers is prohibite
ed. Permission to
re
eproduce in print is res
stricted to an abstract of
o not more than 300 words;
w illustrations mayy not be copied. The abbstract must contain co
onspicuous acknowled dgment of SPE copyrig ght.

AAbstract
AArps’ determin nistic decline cu
urve analysis has
h been the ind dustry standardd for forecastinng production aand estimating reserves
ddespite its subjeectivity and po
otentially large uncertainty, paarticularly earlly in well life. Probabilistic ddecline curve annalysis
((PDCA) metho ods have been proposed
p to quaantify the signiificant uncertaainty in reservees estimation inn shale gas playys.
HHowever, all thhe published PD DCA methodollogies have beeen tested usingg Arps’ declinee curve model. New decline ccurve
aanalysis (DCA)) models have been developeed for hydrauliccally fractured horizontal wellls. However, tthese models hhave been
pprimarily applieed deterministiically, without quantification n of uncertaintyy.

WWe performed hindcast tests ofo six DCA mo odels presented d in the literatuure (Arps, Arpss with a minim
mum decline, m
modified
AArps, Power Laaw, Stretched Exponential,
E annd Duong) with h the Markov C Chain Monte C Carlo (MCMC)) probabilistic
mmethodology with
w different am mounts of prodduction data. Our
O results show w that the MCM MC probabilistic method reliiably
qquantifies the uncertainty
u in production
p hind
dcasts and cummulative producction estimatess for decline cuurve models deeveloped
ffor shale gas wells. Even withh DCA models based on Arpss’ equations, thhe probabilisticc methodologyy is reasonably well
ccalibrated. Hinddcast uncertain
nty decreases as
a the amount ofo matched pro duction data inncreases, but thhe MCMC probbabilistic
mmethod is reasoonably well callibrated regardless of the amo ount of producttion data matchhed. Furthermoore, the MCMC C
pprobabilistic method yields P505 estimates thaat are more acccurate than deteerministic estimmates at early times.

UUncertainty willl always be prresent in produ

uction forecastss and reserves eestimates and uuncertainty cann be quite largee early in
thhe producing lives
l of shale gas
g reservoirs. Reliable
R quantiification of unccertainty can im
mprove decisioon making in eearly stages
oof production, which
w can lead
d to more efficiient exploitatio
on of these reseervoirs.

Introduction
TThere is consid
derable uncertaainty in produ uction forecastiing and reservves estimation for hydraulicaally fractured, horizontal
sshale gas wellss. Methodolog gies that can reliably quantiffy the uncertai
ainty are needeed to allow beetter decision m making by
ooperators in th
he early devellopment and production
p staages of shale gas reservoirrs. Capen (19776) pointed oout that by
pperforming a probabilistic
p an
nalysis, we cann usually obtaiin better estim
mates than usinng a purely detterministic meethodology.
NNevertheless, probabilistic
p deecline curve meethodologies are
a still not widdely used in the industry, eithher because theey are time
cconsuming or hard
h to implemment.

MMajor sources of uncertainty in shale gas production

p foreecasts and reseerves estimates include compplex flow geom metry, large
vvariability in reservoir
r and completion prroperties from well to well,, and lack of long-term prooduction data. Shale gas
rreservoirs posssess extremely low matrix permeability. For this reason , shale reservooirs require massive hydraullic fracture
trreatments to beecome econommical (Agrawal et al. 2012). There
T are also nnatural fracturees in shale, whiich can play ann important
rrole in formation of hydrauliic fracture geo
ometry and dep pletion of the reservoirs (Gaale et al. 20077). There is alsso possible
ddesorption dyn namics of adso
orbed gas (Meengal and Watttenbarger 201 1). All of these phenomenaa result in com mplex flow
ggeometry, whicch contributes to
t uncertainty in
i production forecasting
f andd reserves estimmation.
2 SPE 161300

Another challenge to production forecasting in shale gas reservoirs is the lack of long-term production data. The history of
drilling horizontal wells with multiple hydraulic fractures in shale gas reservoirs is relatively short, only a few years. Because
of their low permeability, shale gas reservoirs require long times, often years, to reach stabilized (boundary dominated) flow.
To the best of our knowledge, only a small number of hydraulically fractured horizontal shale gas wells have experienced
boundary-dominated flow. Conventional decline curve analysis using Arps’ equations require a b value obtained from
stabilized flow (Arps 1945). Such b values cannot be obtained from analyzing transient flow data (Fetkovich et al. 1996).
Despite the lack of a stabilized flow period, many analysts use Arps’ equations coupled with a minimum terminal decline to
estimate reserves and forecast production in shale gas wells (Lee and Sidle 2010).

Several deterministic decline curve models have been developed to work specifically in shale gas reservoirs and address their
unique reservoir and completion challenges, particularly the extensive transient flow period (Clark et al. 2011; Duong 2011;
Ilk et al. 2008; Valko and Lee 2010). Nevertheless, these models by themselves do not quantify the significant uncertainty in
production forecasts and reserves estimates.

PDCA has been used to quantify the uncertainty in production forecasts and reserves estimates in conventional reservoirs
(Cheng et al. 2010; Jochen and Spivey 1996). However, the use of probabilistic methods in hydraulically fractured horizontal
wells in shale gas reservoirs is limited (Gong et al. 2011). We choose to use the Gong et al. (2011) MCMC method because,
compared to Cheng’s (2010) method, the MCMC is faster and provides narrower uncertainty ranges without sacrificing
calibration. Moreover, the MCMC is easier to adapt to new information. The MCMC can allow the integration of other types
of information by modifying the prior distribution.

The MCMC methodology introduced by Gong et al. (2011) has undergone only one test on unconventional wells. The
authors tested the methodology using Barnett shale wells with 80 to 100 months of production data. Most producing wells do
not have long production histories. Hence, we would like to evaluate the calibration of the MCMC method with a limited
amount of data.

The MCMC methodology presented by Gong et al. (2011) was based on Arps’ equations. We believe that Arps’ equations
may not be ideal since new DCAs have been developed for unconventional reservoirs. Further study is needed to evaluate if
other specifically designed shale DCA models can improve the calibration and better quantify the uncertainty in shale gas
reservoirs. Therefore, the objective of this research is to identify the decline curve models that, when coupled with the
MCMC probabilistic methodology; reliably quantify the uncertainty in production forecasts in hydraulically fractured
horizontal shale gas wells, regardless of the stage of depletion.

The large uncertainty in forecasting shale gas production and estimating reserves is a barrier to optimal exploitation of shale
gas resources. Failure to quantify uncertainty in production forecasts and reserves estimates can result in suboptimal
development and economic performance of shale gas assets. Forecast uncertainty may be quite large when limited production
data are available and the uncertainty should decrease as the amount of monthly production data increases. A well calibrated
methodology should be able to reliably quantify uncertainty regardless of the stage of depletion.

In the remainder of the paper we first briefly introduced the elements of the MCMC methodology and describe the decline
curve models employed. We then present the results of a case study in the Barnett shale in which we determine the quality of
calibration of the various different decline curve models in combination with the MCMC probabilistic method.

Bayes’ Theorem and Markov Chain Monte Carlo

The PDCA using Markov Chain Monte Carlo (MCMC) was developed by Gong et al. (2011) to quantify the uncertainty of
production forecasts and reserves estimates. The MCMC method is based on Bayes’ theorem (Eq. 1).
|
| ......................................................................................................................................................................................... (1)
|

In this work, we followed the work of Gong et al. (2011) by allowing θj to be a candidate of the decline curve parameters (δi
for i = 1 to 3 or 4 depending on the DCA model) while y is the historical production data. The objective of a Bayesian study
is to obtain the posterior distribution; in our case we want the posterior distribution of DCA parameters given the production
data.

Bayes’ theorem has three important parameters: (1) the prior distribution of parameters, π(θ), (2) the likelihood function
f(y|θ), and (3) the posterior distribution π(θ|y). The prior distribution is the probability distribution of parameters before any
data have been observed. In our case this is the distribution of DCA parameters before any production data have been
analyzed. The likelihood function is a probability density function of the historical production data assuming the decline
curve parameters are the true parameters. The posterior distribution is the distribution of the unknown parameters after all the
SPE 161300 3

observed data have been considered. From the posterior distribution of DCA parameters, we can calculate the estimated
ultimate recovery (EUR) or reserves. Further details regarding the formal definition of these distributions can be found in
Gong et al. (2011).

Solving for the posterior distribution is difficult when the denominator of Eq. 1 is non-integrable. MCMC simulation is a
technique to estimate the posterior distribution when we face the non-integrable problem. MCMC methods are algorithms for
sampling from probability distributions based on constructing a Markov chain that has the desired posterior distribution.

Gong et al. (2011) used a Metropolis algorithm for MCMC sampling. Since the posterior distribution is unknown, the
samples are drawn from a proposal distribution. The proposal distribution is not necessarily related to what we know before
we draw the samples; it is simply a starting point. The Markov chain should be independent of the proposal distribution,
which means the Markov chains generated using different proposal distributions should converge to the same posterior
distribution. The proposal distribution should not be confused with the prior distribution. The prior distribution is what we
believe the distribution should be before any additional information, and it is part of the calculation of the posterior
distribution. In other words, a proposal distribution is just a starting distribution to sample from, while the prior distribution is
part of the posterior distribution calculation and is more important. Properties of the proposal distribution can be found in
Gong et al. (2011). The standard deviations of the proposal distribution needed to be chosen for each DCA to allow for good
mixing in the MCMC simulation. Good mixing is required so the Markov chain does not remain at one value for too many
iterations.

The prior distribution used by Gong et al. (2011) was an uninformative uniform distribution. The uninformative prior was a
uniform distribution for all the decline curve parameters.

Gong et al. (2011) used a likelihood function that is based on the quality of the fit provided by candidate decline curve
parameters. They assumed that the difference between the log of the actual production and the log of the estimated
production from a DCA model, ε, follows a standard normal distribution (Eq. 2)

| … exp .......................................................................................................................................... (2)

We added a white noise to the production data to model the inherent error in the production data. The white noise was added
to the standard deviation from the residuals as a constant.

Barnett Shale Case Study: 197 Horizontal Gas Wells

We evaluated the performance and benefits of using the MCMC probabilistic method in a sample of Barnett shale gas wells.
We collected public production data from 197 horizontal wells from the Barnett shale. We used the Barnett shale because it is
one of the longest-producing shale plays. The chosen wells were active gas wells from Denton, Tarrant and Wise counties in
Texas. We excluded production data from wells that had undergone an obvious stimulation/recompletion treatment. In wells
in which the stimulation process appeared to be performed near the beginning of the known history, we shifted the starting
date to model the dominant decline trend. With these exclusions, the wells had useable production data ranging from 59 to
119 months, with an average of 80 months.

The analysis was performed using hindcasts. In a hindcast study a user-defined portion of the known history is used as input
to the model (i.e., is matched) and the remainder of the known history is considered to be “future” production and is
compared to the model prediction to the same point in time (i.e., hindcast). We performed six hindcasts by increasing the
amount of production data analyzed from 6 to 36 months using 6-month steps.

We tested three DCA models based on Arps’ equations: Arps, Arps with a minimum decline of 5%, and modified Arps. Arps
with a minimum decline is an extension of conventional Arps’ DCA method. When the instantaneous hyperbolic decline
reaches 5% (commonly used in the industry) in the forecasted production, the model switches from its hyperbolic decline to
an exponential decline of 5%. Modified Arps’ equations include a fourth parameter defined as the time in which the decline
goes to exponential. At the specified time (which is based on the quality of the fit), the model switches from hyperbolic
decline and follows an exponential decline using the instantaneous decline right before the switch.

We also tested three DCA models recently presented in the literature for shale gas resources: Power Law (Ilk et al. 2008),
Stretched Exponential (Valko and Lee 2010), and Duong’s rate-decline analysis for fracture-dominated shale gas reservoirs
(Duong 2011).

We followed the steps proposed by Gong et al. (2011) for using MCMC as the probabilistic method. Hence, we used 1000
iterations for the Markov Chain. The MCMC requires that a proposal distribution be specified for each of the decline curve
4 S
SPE 161300

pparameters useed in each of the DCA mo

odels. We also
o defined our prior distribuutions for declline curve parameters as
uuninformative uniform
u distrib
butions.

TThe MCMC is conducted ind dependently forr each well. Thhe individual wwells’ P90, P50, and P10 estimates are determmined from
thhe distribution
n of each well’s cumulative production
p at th
he end of the hhindcast periodd, based on thee posterior disttribution of
DDCA parameters. However, the set of wellls must be anallyzed to determ mine if the metthod is probabiilistically well calibrated.
WWe believe there is likely corrrelation amongg the wells. Neevertheless, forr simplicity wee assumed thatt the wells werre perfectly
ccorrelated. To calculate the P90, P50, and P10 values forr the set of weells, we addedd the individuual-well P90, P50, and P10
eestimates, respeectively, assum
ming the wells’ individual estiimates are perffectly correlateed (r=1).

WWe used a couuple of tools to assess calibraation of the pro

obabilistic metthodology. Onee was the coveerage rate (CR R), which is
ssimply the num
mber of wells inn which the tru ue production falls within thhe P90-P10 rangee divided by thhe total numbeer of wells.
FFor a well caliibrated method
dology generatting 80% C.I.ss, approximateely 80% percennt of our wells would brackket the true
ccumulative production withinn its C.I.

AAnother tool ussed to assess calibration

c of thhe probabilistic results was a calibration cuurve (Fig. 1). TThe x-axis reppresents the
pprobability assoociated with eaach forecast, such
s as the P900, P50, or P10 esstimates. On thhe y-axis, we pplot the fractioon of wells
thhat comply wiith the definitio
on. For examplle, the well acttual productionn at the end off hindcast shouuld be greater thhan the P10
eestimate for appproximately 10% of the wells being analy yzed. Using thhis definition, w we determinedd the fraction oof wells in
wwhich the actuaal production exceeded
e the P10 estimate an nd plotted this value on the yy-axis. We repeeated this for tthe P50 and
thhe P90 values.

A methodology y is well calibrrated if, for all the estimates assigned

a the saame probability
ty, the proportiion correct is eequal to the
pprobability assiigned (Fig. 1). For high pro obabilities (abo
ove P50), underrconfidence occcurs when thhe probability aassigned is
loower than the proportion corrrect and overconfidence occu urs when probaability assigneed is greater thaan the proportiion correct.
OOn the other haand, for low prrobabilities (below P50), undeerconfidence occcurs when thee probability asssigned is highher than the
pproportion corrrect and overcconfidence occurs when pro obability assiggned is lower than the proportion correcct for high
n other words, being overcon
pprobabilities. In nfident means we
w have narrow wer ranges for our productionn forecasts.

Fig. 1—Calibration
1 plot. Well calib
brated results sh
hould lie on thee unit-slope linee.

WWe evaluated the

t performancce of three decline curve mod dels based on A
Arps’ equationns (Eqs. 3 andd 4). The first m
model used
AArps’ equations with no mod difications. Forr the second mo
odel we used AArps’ equationns with a minimmum decline oof 5%. And
ffor the third mo
odel we used modified
m Arps.

0 ...................................... ......................................................................... (3)

exp 0 ............................ ......................................................................... (4)

W mance of threee decline curvee models deve loped for analyysis of shale ggas reservoirs. The fourth
We also evaluaated the perform
mmodel is the Po
ower Law equaation (Ilk et al. 2008) (Eq. 5). The fifth moddel is the stretcched exponentiial model (SEP
PD) (Valko
aand Lee 2010) (Eq. 6). The siixth model is Duong’s
D (2011)) model (Eq. 77).

................................................................... ......................................................................... (5)

................................................................................... ......................................................................... (6)
SPE 161300 5

1 ................................................................................................................................... (7)
We defined the prior distributions for decline curve parameters as independent, uninformative uniform distributions. The
bounds for these uniform prior distributions (Table 1) were also used in the likelihood function; no values outside these
bounds were considered. The bounds for all the variables are wide enough so any reasonable combination of decline curve
parameters can be considered.
TABLE 1—BOUNDS FOR DECLINE
CURVE MODEL EQUATIONS
DCA Parameter (δ) Upper bound Lower bound

Arps’ Equations
qi, Mcf/D 1,000,000 0.01
Di, 1/year 50 0.1
b 2 0
Arps’ with 5% Min Decline
qi, Mcf/D 1,000,000 0.01
Di, 1/year 50 0.1
b 2 0
Modified Arps’ Equations
qi, Mcf/D 1,000,000 0.01
Di, 1/year 50 0.1
b 2 0
T0, months 10,000 3
Power Law Model
qi, Mcf/D 1,000,000 0.01
Dhat 10 0.001
D∞ 1 1 Exp-09
n 2 0.001
Stretched Exponential Model
qi, Mcf/D 1,000,000 0.01
η 5 0.01
τ 10 0.15
Duong's Model
qi, Mcf/D 1,000,000 0.01
a 5 0.5
m 2 0.5
In the Metropolis algorithm, the proposal distribution, given the previous step candidates of decline curve parameters δi (for i
= 1 to 3 or 4, depending on the decline curve model) is δi,proposal ~ N(δi,t-1, σi). The only unknowns are the standard deviations
for each of the decline curve parameters. Hence, the MCMC requires that a proposal distribution be specified for each of the
decline curve parameters used in each of the DCA models. The standards deviations were chosen to obtain good mixing for
the MCMC simulation (Table 2).
TABLE 2— PROPOSAL DISTRIBUTION
FOR EACH DCA MODEL
DCA Parameter (δ) σ

Arps’ Equations
Ln (qi) 0.2
Ln (Di) 0.4
b 0.2
Arps’ with 5% Min Decline
Ln (qi) 0.2
Ln (Di) 0.4
b 0.2
Modified Arps’ Equations
Ln (qi) 0.2
Ln (Di) 0.4
b 0.2
T 1
Power Law Model
Ln (qi) 0.2
Ln (Dhat) 0.4
Ln (D∞) 0.2
n 0.4
Stretched Exponential Model
Ln (qi) 0.2
Ln(η) 0.4
Ln(τ) 0.2
Duong's Model
Ln (qi) 0.2
a 0.2
m 0.2
6 SPE 161300

Results and Discussion

Our results show that the hindcast cumulative production P90 to P10 range for the set of wells narrows as the amount of
production data matched increases (Figs. 2a to 2f). In other words, uncertainty in the cumulative production at the end of the
hindcast period decreases over time, as expected. In general, we observe that the results are biased if 18 months or less of
production are matched. Bias is the difference between the actual cumulative production and the mean value. As more
production data are matched, the results become less biased. We also observe that for most of the models (modified Arps,
Power Law, SEPD, and Duong’s Model) the P50 estimate for the set of wells is more accurate than the deterministic estimate
when limited production data are available to be matched (less than 18 months).

We observe that narrower C.I.s are produced with Arps’ equations (Fig. 2a to Fig. 2c). However, the results for the Arps
models tend to be more biased if one year or less of production is matched. For the three models based on Arps’ equations the
P50 underestimated the actual production for the set of wells; i.e., the models were pessimistic. We emphasize that the
cumulative production is predicted at the end of known production. Hence, the results for Arps and Arps with a minimum
decline of 5% are fairly similar.

The C.I. for the Power Law model is the second widest (Fig. 2d). However, the results are less biased than Arps’ results at 6
months. We observe that the Power Law model P50 underestimates true production for the set of wells if one year of
production is matched, while it overestimates true production for the rest of the hindcasts. This is the only method that
crosses; all the other methods either consistently underestimate or overestimate. We did not plot the results for 6 months
using the SEPD model because the model yielded unrealistic results for this limited amount of data. Nevertheless, the SEPD
model provided the least biased results overall (Fig. 2e). The SEPD model P50 slightly overestimated actual production for the
set of wells; i.e., the results were optimistic. Duong’s model yields a wider C.I. (Fig. 2d), and it required the most production
data to produce relatively unbiased results. The Duong model P50 significantly overestimated actual production for the set of
wells; i.e., the results were optimistic. We believe that the Duong model P50 overestimates production for the set of wells
because it assumes long-term linear flow for all the wells analyzed, even though some wells may deviate from that trend.

In general, the hindcasts are reasonably well calibrated (i.e., close to the unit-slope line), particularly the P10 and P90 values
(Fig. 3a to 3f). Most of the deviations from the unit-slope line occur in the Arps’ P50 values at early times. Calibration
improves when more production data is matched in the hindcasts. Except for the Duong model, the curves lie at or above the
unit-slope line, indicating the P10, P50 and P90 estimates are too low. Overall, the coverage rate ranges from 64% to 85% for
the 80% C.I.s, with most hindcasts in the 77-82% range.

With the Arps, Arps with minimum decline, and modified Arps models, more production data is required to be well
calibrated compared to the other models (Fig. 3a to 3f). This is likely related to having smaller C.I.s, as seen in Figs. 2a to 2c.
The P10, P50 and P90 estimates all lie on or above the unit-slope line. Thus, the hindcast production estimates should be greater
to be well calibrated. Coverage rates are near 80% for the Arps’ models. This means that approximately 80% of the P10-P90
ranges bracket the actual production.

The hindcasts appear to be better calibrated for the Power Law model, although coverage rates are below 70% for 6 and 12
months of production data matched and are no greater than 74% for more production data (Fig. 3d). This indicates that
coverage rate alone is not sufficient to measure the quality of the probabilistic hindcast. With the SEPD model, 12 months of
production data were required for the model to be stable (Fig. 3e). The coverage rate ranges from 71% to 78% for matches of
12 to 26 months of production data. The SEPD exhibits the best calibration overall, and is well calibrated for all production
data periods matched. The Duong model has the best calibration and coverage rate with 6 months of production data matched
(Fig. 3f), and calibration and coverage rate decrease with more production data matched. Compared to the other models, the
Duong model predictions are consistently optimistic; i.e., the estimates should be smaller to be well calibrated.

The different models have different assumptions, which affect their performance in shale gas reservoirs. In general, when
limited production data are available (6–12 months), models that were not developed for shale gas resources, such as Arps,
tend to underestimate future production. This is because with little data, the steep early declines are fit best with either
exponential declines or hyperbolic declines with relatively low b values, which do not capture the subsequent flattening of
the decline curve. The opposite occurs for the Duong model. The assumption of long-term linear tends to overestimate future
production, at least for the well set examined in this study. We believe that both model types can be improved with use the
use of different prior distributions.
S
SPE 161300 7

(a) (b)

(c) (d)

(e) (f)
F
Fig. 2—MCMC C and a) Arps, b)
b Arps with a 5%5 min decline,, c) modified Arrps, d) Power LLaw, e) Stretcheed Exponential,, f) Duong’s
neral, 80% C.I decrease in sizee as the amountt of production analyzed increases. Also, the results are biassed if less
model. In gen
han 18 months of
th o production data are availab le to match in tthe hindcast.
8 S
SPE 161300

(a)) (b)

(c)) (d)

(e)) (f)

Fig. 3— MC
CMC and a) Arrps, b) Arps witth a 5% min deccline, c) modifieed Arps, d) Pow
wer Law, e) Streetched Exponen
ntial, f)
Duong’s
D model.. In general, ca
alibration is enh
hanced over tim
me using MCMC C as probabilisttic method.

OOverall, there seems

s to be a correspondence
c e between size of the confidennce intervals an and quality of ccalibration. Thee three
AArps models tend to have smaaller confidencce intervals (Fig gs. 2a-2c) and,, correspondinggly, do not apppear to be as well
ccalibrated (Figss. 3a-3c), in spiite of the coverrage rates nearr 80%. The Pow wer Law and D Duong models hhave larger connfidence
inntervals (Figs. 2d and 2f) and d, correspondin ngly, appear to be better calibbrated (Figs. 3dd and 3f), desppite the lower ccoverage
rrates. The SEPD D model appeaars to be somew what of an exceeption to this; iit has small coonfidence intervvals (Fig. 2e) bbut is also
vvery well calibrrated (Fig. 3e).. The primary disadvantage
d of
o the SEPD moodel is that it ddid not work w well with only 6 months
oof production data
d to match.

WWe also note th

hat there does not
n appear to be a correspond dence between visual quality of the calibratiion and the covverage
rrate. Models that appear to bee better calibratted visually (Po
ower Law, SEP PD and Duongg) tend to have coverage ratess that
ddeviate more frrom 80% than models
m that appear to be poorer calibrated vvisually (the A Arps models). T These results coould be
immproved by inncorporating qu uantitative meaasures of the qu uality of calibraation rather thaan relying on vvisual interprettation.
SPE 161300 9

We emphasize that, despite the comparisons between models and qualifications offered, all of the DCA models are
reasonably well calibrated when used in conjunction with the MCMC probabilistic methodology, which we believe is
remarkable given the small amount of production data that is being analyzed in these cases. We believe that, with further
work, the poorer calibration at early production stages (6-12 months of production) for some of the models can be improved
even further.

Finally, we remind the reader that results have been presented for hindcasts only, i.e., predictions to 59-119 months, and not
for long-term forecasts and reserves estimates. While we believe it is reasonable to infer that the probabilistic methodology
will also perform well for long-term forecasts and reserves estimates, this is not guaranteed and cannot be demonstrated until
long-term performance data are available in shale gas reservoirs. While uncertainty will always be present in any production
forecast and reserves estimate, and possibly quite large early in the producing life, reliable quantification of the uncertainty
should improve development decision making by operators.

Conclusions
 The Markov Chain Monte Carlo probabilistic methodology combined with decline curve models developed for
shale gas resources reliably quantifies the uncertainty in production hindcasts to 59-119 months for hydraulically
fractured horizontal shale gas wells in the Barnett shale. Even with DCA models based on Arps’ equations, the
probabilistic methodology is reasonably well calibrated.
 Forecast uncertainty decreases as the amount of monthly production data available for matching increases. The MCMC
probabilistic methodology was reasonably well calibrated regardless of the amount of monthly production data.
However, the hindcast calibration is best with 18 or more months of production data.
 For the modified Arps, Power Law, SEPD and Duong models, the MCMC P50 estimate for the well group is more
accurate than the deterministic estimate at early times.

Nomenclature
a Duong intercept constant, 1/day
b Arps decline exponent, dimensionless
Di Arps initial decline rate, 1/year
D∞ Power Law decline at “infinite time” constant, 1/day
Dhat Power Law decline constant, 1/day
m Duong slope
n Power Law time exponent, dimensionless
P10 Value at confidence level 10%
P50 Value at confidence level 50%
P90 Value at confidence level 90%
qi Initial gas rate, Mcf/D
y Monthly production, Mcf
T0 Modified Arps time go into exponential decline, months
δ DCA parameter
ε Residual of log actual production and log estimated from DCA
θ Parameter acting as random variable
θj Parameter of step j in MCMC
θproposal Parameter drawn from proposal distribution
η Exponent parameter or SEPD model, dimensionless
τ Characteristic time parameter for SPED model, month
σ Sample variance from best fit
σj Sample variance from step j in MCMC
σproposal Sample variance from proposal distribution

References
Agrawal, A., Wei, Y., and Holditch, S. 2012. A Technical and Economic Study of Completion Techniques in Five Emerging
Us Gas Shales: A Woodford Shale Example. SPE Drilling & Completion 27 (1): pp. 39-49. DOI: 10.2118/135396-
pa
Arps, J.J. 1945. Analysis of Decline Curves. AIME 160: 228-247.
Capen, E.C. 1976. The Difficulty of Assessing Uncertainty (Includes Associated Papers 6422 and 6423 and 6424 and 6425 ).
SPE Journal of Petroleum Technology (8): 843-850. DOI: 10.2118/5579-pa
Cheng, Y., Wang, Y., Mcvay, D. et al. 2010. Practical Application of a Probabilistic Approach to Estimate Reserves Using
Production Decline Data. 2 (1): pp. 19-31. DOI: 10.2118/95974-pa
Clark, A.J., Lake, L.W., and Patzek, T.W. 2011. Production Forecasting with Logistic Growth Models. Paper presented at the
10 SPE 161300

SPE Annual Technical Conference and Exhibition, Denver, Colorado, USA. SPE-144790-MS. DOI:
10.2118/144790-ms.
Duong, A.N. 2011. Rate-Decline Analysis for Fracture-Dominated Shale Reservoirs. SPE Reservoir Evaluation &
Engineering (06). DOI: 10.2118/137748-pa
Fetkovich, M.J., Fetkovich, E.J., and Fetkovich, M.D. 1996. Useful Concepts for Decline Curve Forecasting, Reserve
Estimation, and Analysis. SPE Reservoir Engineering 11 (1): 13-22. DOI: 10.2118/28628-pa
Gale, J.F.W., Reed, R.M., and Holder, J. 2007. Natural Fractures in the Barnett Shale and Their Importance for Hydraulic
Fracture Treatments. AAPG Bulletin 91 (4): 603-622. DOI: 10.1306/11010606061
Gong, X., Gonzalez, R., Mcvay, D. et al. 2011. Bayesian Probabilistic Decline Curve Analysis Quantifies Shale Gas
Reserves Uncertainty. Paper presented at the Canadian Unconventional Resources Conference, Alberta, Canada.
SPE 147588. DOI: 10.2118/147588-ms.
Ilk, D., Rushing, J.A., Perego, A.D. et al. 2008. Exponential Vs. Hyperbolic Decline in Tight Gas Sands  ” Understanding
the Origin and Implications for Reserve Estimates Using Arps' Decline Curves. Paper presented at the SPE Annual
Technical Conference and Exhibition, Denver, Colorado, USA. SPE 116731. DOI: 10.2118/116731-ms.
Jochen, V.A. and Spivey, J.P. 1996. Probabilistic Reserves Estimation Using Decline Curve Analysis with the Bootstrap
Method. Paper presented at the SPE Annual Technical Conference and Exhibition, Denver, Colorado. 1996
Copyright 1996, SPE, Inc. 36633-MS.
Lee, W.J. and Sidle, R. 2010. Gas-Reserves Estimation in Resource Plays. (10). DOI: 10.2118/130102-pa
Mengal, S.A. and Wattenbarger, R.A. 2011. Accounting for Adsorbed Gas in Shale Gas Reservoirs. Paper presented at the
SPE Middle East Oil and Gas Show and Conference, Manama, Bahrain. SPE-141085-MS. DOI: 10.2118/141085-
ms.
Valko, P.P. and Lee, W.J. 2010. A Better Way to Forecast Production from Unconventional Gas Wells. Paper presented at
the SPE Annual Technical Conference and Exhibition, Florence, Italy. SPE 134231. DOI: 10.2118/134231-ms.