IIE Transactions (2005) 37, 945956

C IIE
Copyright

ISSN: 0740-817X print / 1545-8830 online
DOI: 10.1080/07408170591008073

A real-options-based analysis for supply chain decisions

HARRIET BLACK NEMBHARD1, , LEYUAN SHI2 and MEHMET AKTAN3
1

Harold and Inge Marcus Department of Industrial and Manufacturing Engineering, The Pennsylvania State University,
University Park, PA 16802, USA
E-mail: hbnembhard@psu.edu
2
Department of Industrial Engineering, University of Wisconsin-Madison, Madison, WI 53706, USA
E-mail: leyuan@engr.wisc.edu
3
Department of Industrial Engineering, Ataturk University, Erzurum, 25240, Turkey
E-mail: maktan2000@yahoo.com

Flexibility allows rms to compete more effectively in a world of short product life cycles, rapid product development, and substantial
demand and/or price uncertainty. We develop a supply chain model in which a manufacturing rm can have the exibility to select
different suppliers, plant locations, and market regions and there can be an implementation time lag for the supply chain operations.
We use a real options approach to estimate the value of exibility and to determine the optimum strategy to manage the exibility
under uncertainty in the currency exchange rate. To price the operational exibility, we develop a Monte Carlo simulation technique
that is able to incorporate a large number of variables into the valuation. We show that without considering time lag impact, the value
of the operational exibility can be signicantly overestimated.

1. Introduction
Manufacturing operations have always been required to
adapt to a changing environment. Some drivers, such as
changes in customer demand and shifts in competition,
have always been present. Other forces (such as higher degrees of product proliferation and a more global marketplace) are relatively recent additions. The domination of
customer demand has also dramatically shifted the product market. Customers require more product diversity and
expect to see new products regularly introduced into the
market. In any case, the change that such forces bring about
typically present signicant challenges for many aspects of
an operation. One strategy has been to increase the exibility of the manufacturing operation. Flexibility allows rms
to compete more effectively in a world of short product life
cycles, rapid product development, and substantial demand
and/or price uncertainty.
In the forward to Real Options (Trigeorgis, 1999), Mason
writes: Flexibility has value. While this statement is obvious at the conceptual level, it is surprisingly subtle at the
applied level. The nancial arena was the original proving
ground for the application of the options-based framework
to the valuation of exibility. More recently, managerial
operating exibility has been likened to nancial options.

Corresponding author

C 2005 IIE
0740-817X

Flexibility is usually acquired at a cost premium; therefore proper valuation of benets is important to justify the
added cost. Without such work rms may under-invest or
make an entirely wrong decision about key technologies
or strategies. Thus, exibility potentially provides a competitive advantage to the rm. The question then becomes:
precisely how valuable is exibility? The goal of real options
valuation is to address this question in an environment of
high volatility.
Nembhard et al. (2000) suggested the need for modeling
the exibility surrounding manufacturing operations using
real options. An important issue in applying real options
to manufacturing enterprises is to incorporate the practical operational concerns of the decision-making into the
existing real options framework (Brach, 2003). Financial
options can be exercised instantly; the trader simply tells
the broker to move the contract and the deal is done. In
the existing body of literature on applying real options, the
common assumption is that a real option can be implemented immediately as well. However, in complex manufacturing operations, this assumption does not truly hold.
Manufacturing operations need some time to be executed.
When a decision is made to exercise a real option, there is
some period of time until the decision can be implemented.
Accordingly, in this paper we consider the implications of
a time lag on real options valuation. The main purposes are
to capture this difference into the framework and to analyze
its effect on the outcome and hence the managerial course

946
of action. To facilitate the work, we adapt current solution
procedures to address this problem. Real options whose values depend on multiple sources of uncertainty, rarely have
closed-form solutions so numerical approximation procedures must be used. Specically, we use a multinomial lattice
approach and we develop a Monte Carlo simulation procedure that can be applied in option valuation analysis when
there is a time lag between the decision and the implementation of the option. We nd that whereas the lattice approach
does make a good approximation, a Monte Carlo simulation procedure can more easily incorporate a large number
of variables into the valuation. Using the proposed procedures, a company will be able to obtain better estimates for
the option value, and to produce optimal decisions considering the effect of a time lag between the decision and
the implementation of the option. We show that without
considering time lag impact, the value of the operational
exibility can be signicantly overestimated.
There has been some research in the nancial engineering
literature that is relevant to this problem. McDonald and
Siegel (1985) presented option pricing techniques to study
the investment problem of a rm that has the option to
shut down or change the level of production. Brennan and
Schwartz (1985) presented a problem where the option was
to open, close, or abandon a mine. Majd and Pindyck (1987)
used options analysis to derive optimal decision rules and
to value investment decisions where construction proceeds
exibly and can be adjusted with the arrival of new information. Hodder and Triantis (1993) presented a general framework for modeling and evaluating investments that involve
having exibility to switch between alternative states of operation. Kogut and Kulatilaka (1994) modeled the operating exibility to shift production between two manufacturing plants located in different countries using a stochastic
dynamic programming model. Huchzermeier and Cohen
(1996) developed a stochastic dynamic programming formulation for the valuation of global manufacturing strategy options with switching costs, where a rm maximizes
its expected, discounted, global, after-tax value through
the exercise of product and supply chain network options
through exploitation of exibility contingent on exchange
rate realizations. Dasu and Li (1997) studied the structure
of the optimal policies for a rm operating plants in different countries. They determined the structure of the optimal
policies for deciding when and by how much to alter the production quantities. Smith and McCardle (1998) presented
how stochastic dynamic programming and option pricing
theory can be protably integrated to evaluate oil properties
where production rates and oil prices both vary stochastically over time and, at any time, the decision-maker may terminate the production or accelerate production by drilling
additional wells. Markov Decision Processes (MDPs) have
been applied to a wide range of stochastic control problems
including inspection-maintenance-replacement systems, inventory management, and economic planning and consumption models. A link between MDPs and options can
be seen through Tsitsiklis work: Tsitsiklis and Van Roy

Nembhard et al.
(1999) developed a theory characterizing optimal stopping
times for discrete-time ergodic Markov processes with discounted rewards; Marbach and Tsitsiklis (2001) proposed
a simulation-based algorithm for optimizing the average
reward in a nite-state Markov reward process that depends on a set of parameters; and Tsitsiklis and Van Roy
(2001) introduced and analyzed a simulation-based approximate dynamic programming method for pricing American
options. Focusing more specically on manufacturing systems, Nembhard et al. (2002) valued real options associated
with the exibility to apply statistical process control charts
to monitor quality in a production process. Nembhard et al.
(2003) studied the option value of being able to switch between the states of producing or outsourcing an item using
Monte Carlo simulation.
In this paper, we extend our research application to value
the real options encompassing the exibility to switch between global manufacturing options for a rm that has
operations in different countries and there may be an implementation time lag for the supply chain operations. The
rm maximizes its expected discounted value through the
exercise of the manufacturing options in an environment
of uncertain exchange rates. The remainder of this paper
is organized as follows. We present the problem formulation in Section 2. Valuation procedures that can be used to
solve the problem, and the proposed solution techniques
for the problem are presented in Section 3. Section 4 gives
the numerical results. We make some concluding remarks
in Section 5.

2. Model formulation for real options with a time lag

We consider a supply chain network model for a global
rm that has suppliers, production plants, and markets in
different countries. Raw material ows from the suppliers to
the plants. The raw material is turned into nished product
in the plants, and then the nished product is transferred to
the market regions in different countries. The rm denes
a set of manufacturing real options, where the suppliers,
plants, and market regions are selected. Each option denes
which suppliers, plants, and markets of the supply chain
network will be used. The strategy of the rm is to switch the
suppliers, plants, and markets according to opportunities
presented when exchange rates change.
We assume that the decision-maker is concerned with increasing manufacturing exibility through this strategy as
well as knowing the impact of making a switching decision
when there must be a time lag prior to its implementation.
For example, when the company wants to switch the supplier, all raw material from the new supplier cannot be received instantly; when a switch is desired for the production
plants, the required amount of production in the new plant
cannot be completed instantly; and when a switch is desired
for the market regions, nal product cannot reach the new
market instantly.

947

Real-options-based analysis for supply chain decisions

A time horizon is determined for the problem, and it is
divided into time intervals of equal length. In each time
interval, the rm will select one of the manufacturing options. If there is a switch between the previous and the current decisions, this results in a switching cost. Each switching cost is dened depending on which suppliers, plants,
and market regions must be changed to exercise the switch.
In our rst analysis, we assume that a switch cannot
be implemented in the time interval that the decision is
given; rather it will be implemented in the next time interval, i.e., with a one period time lag. In our second analysis,
we assume that a portion of the new option will be implemented in the time interval that the decision is given, and
the remaining portion will be implemented in the next time
interval.
We formulate a model for the option valuation problem under a time lag as a stochastic dynamic program.
This model builds upon the structure of Huchzermeier and
Cohen (1996) model. The following notation is used for
modeling the problem:
= time index for t = 0, 1, . . . , T;
= vector of exchange rates for the foreign
countries at time t;
Ot
= option implemented at time t;
Pt (et , Ot1 , Ot ) = prot gained during time interval beginning at time t, if option Ot1 was
implemented at time t 1, option Ot
is implemented at time t, and the vector of currency exchange rates at time
t is et ;
= total expected prot between time t
Vt (et ,Ot1 )
and time T, given that the vector of currency exchange rates is et , and the option implemented at time t 1 is Ot1 ;
= cost of switching from option Ot1 to
WOt1 Ot
option Ot ;
= unit raw material cost of supplier i;
ai
= unit manufacturing cost at plant i;
ci
Ri
= price of the rms output in market region i;
= number of units in the shipment from
xspt
supplier s to plant p during time interval t;
= number of units in the shipment from
xpmt
plant p to market m during time
interval t.
t
et

In each state, the prot is maximized by selecting an

option Ot , given that the rm selected option Ot1 in
the preceding time interval, and the vector of current exchange rates between the home country and foreign countries is et . The value V of the total prot at time t and
state (et , Ot1 ) is dened with the following recursive
equation:
Vt (et , Ot1 ) = max E[Pt (et , Ot1 , Ot )]
Ot

+ er
t Vt+1 (et+1 , Ot ).

(1)

In Equation (1), Ot1 is the option applied in time interval

t1, Ot is the option selected for the time interval t, E[Pt (et ,
Ot1 , Ot )] is the expected prot during time interval t, and
Vt (et , Ot1 ) is the total expected prot from time interval t
to the last time interval.
Huchzermeier and Cohen (1996) applied a version of
Equation (1) to their problem without time lags. We found
that Equation (1) can also be applied to problems with time
lags, but with an important distinction about how it is applied. In Section 3, we will present these application techniques, and we compare the application of Equation (1) to
the two types of problems.
Denoting the cost of switching from option Ot1 to option Ot as WOt1 Ot , and suppressing the time subscript t,
we dene the prot Pt (et , Ot1 , Ot ) as follows:



Pt (et , Ot1 , Ot ) =
(em Rm ep cp )xpm
p

p

es as xsp WOt1 Ot

(2)

where subscripts s, p, and m are the set of suppliers, plants,

and markets included in the option Ot , respectively.
An advanced form of the time lag problem is the one
where a portion of the new option is implemented in the
time interval that the decision is given. Let Q be the portion
implemented in the current time interval, where 0 Q 1.
With a lag of one time interval, and with the portion Q of
the option implemented in the time interval of the decision,
the recursive equation for value V of the total prot at time
interval t is dened as:
Vt (et , Ot1 ) = QPt1 (et1 , Ot1 , Ot )
+ max(1 Q)E[Pt (et , Ot1 , Ot )]
Ot
r
t

+e

Vt+1 (et , Ot ).

(3)

Currency exchange rates between the home country and

the foreign countries are the sources of uncertainty and
volatility in the problem. We assume that the exchange rate
ei between the home countrys currency and the currency of
foreign country i follows geometric Brownian motion as:
dei,t
= i dt + i dzi ,
ei,t

(4)

where i is the drift of the exchange rate changes for foreign country i, i is the volatility of the exchange rate
for foreign country i, dz is a standard Wiener disturbance
term. The expected changes in the exchange rates are set to
i = exp[(r ri )
t], where
t is the length of each time
interval, and r and ri are the risk-free rate of return in the
home country and country i, respectively.
In the next section, we discuss two option valuation methods that can be used to model the state variable movements
for this problem, and we present our procedures that use
these two methods for valuation of the problem with the
feature of implementation time lag.

948

Nembhard et al.

3. Valuing real options with a time lag

The state variables in our problem are the currency exchange rates between the home country and the foreign
countries. In order to model the exchange rate movements
and value the options, we will use multinomial lattice and
Monte Carlo simulation procedures.
If there were no switching costs between the options, each
time interval could be evaluated independent from other
time intervals. Then, we could simply maximize the prot
in each time interval without considering the decisions in
other time intervals. Valuing a real options problem with
switching costs is more difcult since options exercised in
successive time intervals have connections with each other,
and the current decision inuences the later ones. Hence, we
cannot separate the problem into time intervals where the
decisions are independent from one another. When switching costs exist, we need to apply a dynamic programming
approach to maximize the prot.
We will illustrate the numerical valuation procedures using the simple supply chain network shown in Fig. 1. In this
network, there are two different countries. There is one supplier and one production plant in each of the two countries.
At the same time, each of these two countries is a market
region for the nal product. The manufacturing options are
dened based on possible connections in the supply chain
network.
Figure 2 shows a set of 12 manufacturing options which
are dened considering the supply chain network in Fig. 1.
Each option shows the connections between the suppliers,
plants, and market regions. So, for example, the rm may
initially be operating under option 8, where supplier 2 provides materials to plant 2 which sells goods in market 2, and
some time period later decides to switch to option 6, where
supplier 1 provides material to plant 2 for sales in market 2.

Fig. 2. Twelve manufacturing options for the two suppliers, plants,

and markets.

val. If we have two exchange rates, there are ve possible

movements (i.e., for e1 and e2 there are four combinations
of up and down movements as well as the possibility of no
change), each with a probability of occurrence (Kamrad
and Ritchken, 1991).The rate of movement for the currency
exchange rate between the home country and the foreign
country i is given as:
ui = ei

t

Figure 3 shows the effect of decisions in the lattice for two

state variables. There are ve nodes emerging from each
node. The rst element in parenthesis is the state of e1 , and

3.1. The Multinomial lattice procedure

Building upon the binomial lattice approach of Cox et al.
(1979), Kamrad and Ritchken (1991) developed a multinomial lattice procedure for valuing options with multiple
state variables. Assume that the exchange rate ei can move
up with the rate of ui , or move down with the rate of di
such that di = 1/ui , or stay constant at each time inter-

Fig. 1. A supply chain network for two suppliers, plants, and

markets.

Fig. 3. Effect of decisions in the lattice.

949

Real-options-based analysis for supply chain decisions

the second element is the state of e2 . Since we assume that a
decision will be implemented in the next time interval, the
decision given in a node can be implemented in the immediately following nodes. At each node, we must select the best
option that can be implemented in the next time interval.
In order to do that, we select the option that maximizes the
recursive function (see Equation (1)) for the expected prot
in the immediately following ve nodes.
For the dynamic programming formulation, we need to
select the option that yields the largest expected prot at
each group of nodes for every possible previous option.
We begin from the last time interval and go back one time
interval at each iteration. When we reach the rst node,
the optimal solution is determined. The complete set of
decisions maximizing the total expected prot is obtained
by backtracking in the dynamic program.
If there is no time lag to implement the decision, the
recursive equation maximizing the prot Vt (et , Ot1 ) at
time interval t in a multinomial lattice is dened as:
Vt (et , Ot1 ) = max [Pt (et , Ot1 , Ot )
Ot

+ er
t

n

pj Vt+1 (et+1,j , Ot )],

(5)

j=1

where n is the number jumps from each node, Ot1 is the

option in the preceding node at time interval t1, et is the
vector of exchange rates in the current node at time interval
t, and et+1,j is the vector of exchange rates in node order j
of the following n nodes at time interval t+1.
With a time lag of one time interval, the recursive equation maximizing the prot Vt (et ,Ot1 ) at time interval t is
dened as:
n


pj [Pt (et,j , Ot1 , Ot )
Vt (et , Ot1 ) = max
Ot

+e

j=1
r
t

Vt+1 (et+1,j , Ot )].

(6)

where e t is the set of exchange rates in a group of n nodes at

time interval t, et,j is the vector of exchange rates in node
order j of these n nodes at time interval t, and e t+1,j is the
set of exchange rates in the n nodes that follow the node
that has the vector of exchange rates et,j . Hence, option Ot
which is selected at time interval t 1 will be implemented
during time interval t. The selected option Ot maximizes
the total expected prot in the n nodes at time interval t
and the following nodes until the last time interval.
With a lag of one time interval, and with a portion Q of
the option implemented in the time interval of the decision,
the recursive equation maximizing the prot Vt (et ,Ot1 ) at
time interval t is dened as:
Vt (et , Ot1 ) = QPt1 (et1 , Ot1 , Ot )
n


+ max
pj [(1 Q)Pt (et,j , Ot1 , Ot )
Ot

j=1

+ er
t Vt+1 (et+1,j , Ot )],

(7)

where, portion Q of the option Ot which is selected at time

interval t 1 is implemented during time interval t 1,
and the remaining portion (1Q) of the option Ot is implemented during time interval t.
Because of the no-arbitrage principle in the lattice, the
discounted expected value of a state variable for one time
period will be equal to the initial value of that variable, i.e.:
et = er
t

n

pj et+1,j .

j=1

Therefore, apart from a discount factor of a possible switching cost, the expected prot due to the optimum decision
given at time t and implemented with a time lag at t+1,
and the prot due to the optimum decision given at time t
and implemented immediately at time t, are in fact equal.
Hence, the optimum strategy that will be implemented at
time t+1 in the time lag problem, and the resulting prot
value can both be determined by nding the optimum decision for time t in the original lattice without time lags. Prot
at the rst time interval for the time lag problem can be calculated using the given strategy that must be implemented
at the rst time interval.
With lattice methods, calculation effort increases dramatically with an increasing number of state variables, which
makes them impractical and difcult to use when there are
more than two state variables in a model. In Section 3.2,
we propose a Monte Carlo simulation technique, which
simplies valuing exibility in models with multiple state
variables.
3.2. The Monte Carlo simulation technique
Boyle (1977) introduced a Monte Carlo simulation method
for asset pricing of European options. This approach,
however, cannot be used for problems with switching
costs since the problem cannot be treated as a bundle of
European options with different expiration dates. Broadie
and Glasserman (1997) developed a simulation algorithm
for estimating the prices of American-style assets. In order to develop valid error bounds on the true option value,
they introduced two estimators, one biased high and one
biased low, but both asymptotically unbiased as the computational effort increases. These estimators are based on
simulated lattices. The simulated lattices are parameterized
by b, the number of branches per node. State variables are
simulated at a nite number of possible decision points, i.e.,
exercise times.
In this section, we develop a Monte Carlo simulation procedure that can be used to value the problem in Section 3.1
which has nonzero switching costs, multiple decision points
to switch between the options, and time lags between the
decision and the implementation. There are signicant
differences between the American stock option and the
given manufacturing strategy planning problem with supply chain network: a stock option is exercised instantly, but

950

Nembhard et al.

manufacturing options are implemented after the exercise

decision with some time lag; a stock option is a single option, but there are multiple options in the manufacturing
strategy planning problem; the stock option is exercised
once, but there are multiple exercises of manufacturing options; there are switching costs between the manufacturing options; values at successive time periods are summed
up rather than discounting a single exercise value of an
American option.
In our procedure, we use a simulation tree with b branches
per node. On the simulation tree, we apply the following
dynamic programming algorithm to calculate the high estimator. During the backward recursion of the dynamic
program, we use the expected prot of the b branches to
select the best manufacturing option to be implemented
in those branches. In other words, we dene the best decision of a node so that the implementation of this decision at the next time interval gives the maximum average
prot. Since the simulated values for the state variables
are used for calculations, the dynamic program takes advantage of knowledge of the future to overestimate the
option value. This results in an estimate that is biased
high.
In order to nd the low estimator, we use b-1 branches to
determine a policy, and apply that policy in all b branches.
We repeat this process for the b possible combinations of
b-1 branches. Then, we take the average of the prots of
the b combinations. Since the decision is determined by b-1
branches instead of b branches, the expected prot tends
to be biased low. We take average of the high-biased and
low-biased estimates to estimate the true value.

itm =

Fig. 4. Effect of decisions in the simulated tree.

t 1 is i:
gijt (s) = E[hijt (s) + eR fj,t+1 (St+1 ) | St = s],
fit (s) = max {gilt (s)} .
l=1,...,k

The high estimator  is dened recursively by:

1

fiT (ST ),
iT =
b
and



1
[hilt (St ) + eR l,t+1 ] .
it = max
l=1,...,k b
The low estimator
is dened recursively as follows. First let

iT = (1/b) fiT (ST ). Next dene:

b
b
b
1

 n

 n


 n

1

1

R
R

hilt St + e l,t+1 =
hiit St + eR i,t+1 ,
hiit St + e i,t+1 if max

l=1,...,k
b n=1
b1 n=1
b1 n=1

n = m
n = m

b
b
b
1

 n

 n


 n

1

1

R
R

S
+
e
h
S
+
e
hijt St + eR j,t+1 ,
=
h

if
max

ijt
j,t+1
ilt
l,t+1
t
t

l=1,...,k

b n=1
b1 n=1
b 1 n = 1,

n = m

Figure 4 shows the effect of decisions in a two-step simulated tree for b = 3, where et,j is the vector of exchange
rates at time interval t and node order j.
We use the following notation for the procedure:
t
= 0, 1, . . . , T and is the time index;
{St : t = 0, 1, . . . , T} = a risk-neutralized Markov chain
recording all state variables;
= a discount factor for one time
eR
interval.
Also, hijt (s) is the prot in time interval t by exercising option j in state s, if the previously exercised option in time

n = m


for m = 1, . . . , b. Then let it = (1/b) bm=1 itm for t =
0, . . . ,T 1. To simulate the path followed by ei , we approximate Equation (4) as:




i2

t + i
t . (8)
ei (t +
t) = ei (t) exp i
2
Now, we give a example for the application of our procedure. Assume that we only have two manufacturing options
and two decision points. W12 = $4.0 is the switching cost
from option 1 to option 2, and W21 = $3.0 is the switching
cost from option 2 to option 1. Assume that by running
a single simulation run, we have generated the values of

Real-options-based analysis for supply chain decisions

Fig. 5. Prots obtained by option 1 and option 2 at each node.

the state variables at each node in the tree, and using the
generated values, we have calculated the prots for each
node as given in Fig. 5. The values in parentheses are the
prots in the home countrys currency for option 1 and option 2, respectively. For simplicity of the calculations, we
assume that the risk-free rate of return is zero.
Assume that the current strategy is option 1, i.e., the option that is applied in the initial node is option 1. Figure 6
shows the high estimates in the tree, and the options that
yield the high estimates. We begin to work on the tree from
the last time interval. To nd the high estimate, we calculate
the expected prot of b branches to select the best manufacturing option to be implemented in those branches.
As a sample calculation, let us use the top three branches
in the last time interval. If we apply option 1, the average
prot is (80 + 120 + 90)/3 = 96.67. If we apply option 2, the
average prot is (70 + 100 + 110)/3 = 93.33. If the previous
option was option 2, and if we decide to switch to option 1,
then the average prot is 96.67 3 = 93.67 > 93.33. Then,
switching to option 1 gives a larger prot value than keeping

Fig. 6. High estimates and the options selected.

951
option 2. Therefore, whatever the previous option is, we
should select option 1. In Fig. 6, the values (96.67, 93.67)
represent the prot estimates for the previous options 1
and 2, respectively. After completing the calculations for all
branches in the last time interval, we move to the previous
time interval. If we apply option 1, the prot estimate is
[(110 + 96.67) + (90 + 96.67) + (95 + 100)]/3 = 196.11. If
we apply option 2, the prot estimate is [(100 + 93.67) +
(110 + 100) + (90 + 97)]/3 = 196.89. The difference between
the two prot estimates is 196.89 196.11 = 0.78, which is
smaller than W21 . This means that we must continue with
the previous option. Then, for the three time intervals, we
get a total prot estimate of 100 + 196.11 = 296.11, which
is the high estimate.
Figure 7 shows the low estimates and the options that
yield the low estimates. In order to calculate a low estimate,
we use b 1 branches to determine a policy, and we apply
that policy in all b branches. We repeat this process for the
b possible combinations of b 1 branches. Then, we take
the average of prot estimates obtained from these b combinations. As a sample calculation, let us use the top three
branches in the last time interval. If we apply option 1 to
the rst and second branches, the prot estimate is (80 +
120)/2 = 100. If we apply option 2 to those branches, the
prot estimate is (70 + 100)/2 = 85. Since the difference
between the prot estimates is 100 85 = 15 >W21 , we
select option 1 whatever is the previous option. If we apply
option 1 to the rst and third branches, the prot estimate
is (80 + 90)/2 = 85. If we apply option 2 to those branches,
the prot estimate is (70 + 110)/2 = 90. We see that the difference between the prot estimates is 90 85 = 5 > W12 .
Thus, if the previous option is option 1, and if we switch to
option 2, the estimated prot is 90 4 = 86 > 85. Then,
whatever is the previous option, we select option 2. If
we apply option 1 to the second and third branches, the
average prot is (120 + 90)/2 = 105. If we apply option 2 to
those branches, the average prot is (100 + 110)/2 = 105. We

952

Nembhard et al.

Fig. 7. Low estimates and the options selected.

see that the difference between the average options is

105 105 = 0, which is smaller than W12 and W21 . Hence,
we continue to apply the previous option.
Now, we have three decisions from the three combinations we have tried, and those decisions are apply option
1, apply option 2, and apply the previous option. If
the previous option is option 1, the average prot by applying these three decisions to the three branches is [{(80 +
120 + 90)/3} + {(70 + 100 + 110)/3 4} + {(80 + 120 +
90)/3}]/3 = 94.22. If the previous option is option 2, the
average prot by applying the three decisions to the three
branches is [{(80 + 120 + 90)/3 3} + {(70 + 100 +
110)/3} + {(70 + 100 + 110)/3}]/3 = 93.44.
After completing the calculations for the other branches
in the last time interval, we move to the previous time interval. If we apply option 1 to the rst and second branches,
the prot estimate is [(110 + 94.22) + (90 + 96.44)]/2 =
195.33. If we apply option 2 to those branches, the prot
estimate is [(100 + 93.44) + (110 + 95.78)]/2 = 199.61.
Since the difference between the prot estimates is 199.61
195.33 = 4.28 >W21 , we select option 1 whatever is the previous option. After completing the other two combinations
of branches, we obtain these three decisions: apply option
2, apply option 2, and apply option 1. Applying these
decisions to the three branches, we obtain the prot estimates of 192.63 and 194.30 for previous options 1 and 2,
respectively. Since the initial option is option 1, the total
low estimate for the three time intervals is 100 + 192.63 =
292.63. The Monte Carlo simulation estimate for the prot
is (296.11 + 292.63)/2= 294.37, which is the average of the
high and low estimates.

4. Numerical results
In this section, we present the results of the lattice and
Monte Carlo simulation approaches using a case example.
A rm wants to maximize its prot from the operations on

the supply chain network given in Fig. 1. The total time

horizon for the problem is 1 year. Four time intervals are
dened, which implies that each time interval is 3 months.
In each time interval, the rm will exercise one of the 12
manufacturing options given in Fig. 2. Parameters used in
the problem are as follows:
r
r1
r2
a1
a2
c1
c2
xsp
xpm
R1
R2
D1
D2
1
2

=
=
=
=
=
=
=
=
=
=
=
=
=
=
=
=

5% (risk-free interest rate in the home country);

4% (risk-free interest rate in country 1);
2% (risk-free interest rate in country 2);
$1.1 (unit raw material cost of supplier 1);
$1 (unit raw material cost of supplier 1);
$1.05 (unit manufacturing cost at plant 1);
$1 (unit manufacturing cost at plant 2);
amount of shipment from supplier s to plant p;
amount of shipment from plant p to market m;
$2.3 (price of the rms output in market region 1);
$2.2 (price of the rms output in market region 2);
1100/quarter (demand in market region 1);
1000/quarter (demand in market region 2);
0.2 (volatility for exchange rate of country 1);
0.133 (volatility for exchange rate of country 2);
0.3 (correlation for the exchange rates).

The matrix of switching costs WO O between the 12 options

is shown in Fig. 8.
Figure 9 shows the expected prots obtained by the lattice
procedure for no switching and costly switching. If there is
no switching, the same option will be used at all time intervals. However, when switching is possible, the best option
will be selected at each time interval considering the switching costs. Then, value of exibility because of being able to
switch the manufacturing strategy is the difference between
the expected prots of costly switching and no switching.
The upper line in Fig. 9 shows the expected prots when
switching between options is possible and the decision can
be implemented in the current time interval. The middle
line shows the expected prots when switching between options is possible but the decision can be implemented in the

Real-options-based analysis for supply chain decisions

953

Fig. 8. The matrix of switching costs.

next time interval. The lower line shows the expected prots
when there is no switching between the options. The horizontal axis shows which manufacturing option has been
used prior to the rst time interval in the problem.
If the decisions can only be implemented with a time lag,
and if we ignore this limitation and solve the problem as
if there is no time lag for the implementation, then we will
overestimate the value of exibility. The amount of overestimation is the distance between the upper line and the middle
line in Fig. 9. When a decision cannot be implemented in
the rst time interval immediately, we must implement the
strategy that was selected before the decision time horizon,
whereas a strategy can be selected and implemented immediately in the rst time interval when there is no time lag.
Therefore, if the strategy selected before the rst time interval is not the optimal strategy for the rst time interval,
then there is a prot loss because of implementation time
lag. At the same time, the best strategy can be selected and
implemented at the last time interval when there is no time
lag, but a strategy which was selected earlier without observing the last state will be implemented in the last time

Fig. 9. Expected prots with and without switching.

interval when there is time lag. This is the other source of

prot loss when there is implementation time lag.
When there is no implementation time lag for the options, the estimated value of manufacturing exibility is the
distance between the upper and the lower lines in Fig. 9.
When there is an implementation time lag, the estimated
value of the manufacturing exibility is the distance between the middle and lower lines in Fig. 9.
Figure 10 shows the expected prots for Q = 1, Q = 0.5,
and Q = 0. A value of Q = 1 means that the new option
can be fully implemented in the current time interval (i.e.,
there is no time lag between the decision and the implementation), and we obtain the expected prots shown in
the upper line for Q = 1. A value of Q = 0.5 means that
50% of the new option is implemented in the current time
interval, and we obtain the expected prots shown in the
middle line. A value of Q = 0 means that the new option
cannot be implemented in the current time interval (i.e.,
the switch can be implemented in the next time interval, so
there is a time lag of one time interval) and we obtain the
expected prots shown in the lower line.

954

Nembhard et al.

Fig. 10. Expected prots for different time lags.

With an increasing number of time intervals, the decrease

in the expected prot due to implementation time lag will
become relatively smaller against the total expected prot.
Since we considered four time intervals in our analysis, the
ratio of the expected prot loss due to implementation time
lag to the total expected prot is large. The ratio of prot
loss to the total prot will become smaller with increasing
time intervals, because the difference is due solely to the
rst and the last time intervals.
Again we note that it is important to appreciate the effects
of time lag implementation and also the degree to which a
part of the decision must be deferred. We can make a more
conservative statement about the expected prot assuming
a one-period lag. As we are able to rene our understanding
of the true portion of the option that can be implemented
immediately, we can obtain a better statement of the expected prot. Applying the proposed Monte Carlo simulation procedure on the same supply chain network problem,

Fig. 11. Comparison of the lattice and simulation estimates.

we obtained the simulation results given in Fig. 11. For each

initial option, we made 1000 simulation runs.
Lattice and simulation estimates for the expected prots,
and the percent difference between the estimates are given
in Table 1. We see that the Monte Carlo simulation estimates are close to the lattice estimates, which implies that
our Monte Carlo simulation procedure yields close results
to the analytical solution for this problem. This outcome of
similarity between the procedures has been shown in numerous studies (for example those of Boyle (1988), Boyle et al.
(1989), Kamrad and Ritchken (1991), Amin and Khanna
(1994), and Nembhard et al. (2002)).
Many manufacturing system models contain multiple
state variables. If there are more than two state variables
in a problem, lattice techniques become difcult to apply.
If we want to apply the lattice approach to such problems,
the structure of the lattice becomes too complex and the
number of possible jumps at each node becomes too large.

955

Real-options-based analysis for supply chain decisions

Table 1. Lattice and simulation estimates for valuing the switching option
Initial option

Lattice
Simulation
Percentage difference

10

11

12

1587
1588
0.11

1740
1701
2.23

1795
1785
0.57

1949
1905
2.28

1472
1469
0.21

1620
1585
2.17

1670
1661
0.58

1819
1778
2.27

1735
1735
0.03

1939
1897
2.16

2044
2035
0.43

2248
2230
0.78

However, the proposed Monte Carlo simulation procedure

does not require a different tree structure when the number of state variables increases. The same tree structure can
be used for any number of state variables; the only difference lies in the number of random numbers generated. The
main advantage of the proposed Monte Carlo simulation
approach is its simple application to real options problems
with three or more state variables.

5. Conclusions
Flexibility allows rms to compete more effectively in an
environment of uncertainty. A real options approach estimates the value of exibility and is used to determine the
optimum strategy to manage the exibility. In this study,
we incorporated the practical operational concerns of
decision-making in manufacturing enterprises into the existing real options framework by considering the case where
there may be an implementation time lag. The main purposes were to capture the difference between immediate implementation and the time lag into the framework, and to
analyze the effect on the outcome and hence the managerial
course of action. We adapted current solution procedures
to address this problem. We found that when there is a
time lag, the option value is smaller than the case of immediate implementation. However, the ratio of the option
value decrease to the total option value becomes smaller
when the number of time intervals in the analysis increases.
Therefore, the necessity to analyze the effect of the implementation at the time lag decreases with an increasing number of decision points in the analysis time horizon.
The real options valuation procedures discussed in this
paper give decision-makers a way to choose the appropriate
manufacturing enterprise strategy based on an integrated
view of the market dynamics. We have presented the application of the procedures to a supply chain network with
uncertain exchange rates, where costly switching decisions
for the suppliers, production plants, and market regions require a time lag to be implemented. Overall, the manufacturing enterprise maximizes its expected discounted prot
through effective decisions to exercise real manufacturing
options across the supply chain. In such problems, there
are multiple options. At each time interval, switching between options is possible, and each switch results in a switching cost. A comparison of the results for the Monte Carlo
simulation procedure and the lattice procedure has shown

that the proposed Monte Carlo simulation procedure yields

close estimates for the true option value. Whereas the lattice approach is an efcient tool for cases containing one
or two state variables, a Monte Carlo simulation procedure
can more easily incorporate a large number of variables into
the valuation.

Acknowledgement
This research has been partially supported by the National
Science Foundation under award #135-8459.

References
Amin, K.I. and Khanna, A. (1994) Convergence of American option values from discrete to continuous-time nancial models. Mathematical
Finance, 4, 289304.
Boyle, P.P. (1977) Options: a Monte Carlo approach. Journal of Financial
Economics, 4, 323338.
Boyle, P.P. (1988) A lattice framework for option pricing with two state
variables. Journal of Financial and Quantitative Analysis, 23(1), 112.
Boyle, P.P., Evnine, J. and Gibbs, S. (1989) Numerical evaluation of multivariate contingent claims. The Review of Financial Studies, 2(2),
241250.
Brach, M. (2003) Real Options in Practice, Wiley, Hoboken, NJ.
Brennan, M.J. and Schwartz, E.S. (1985) Evaluating natural resource
investments. Journal of Business, 58(2), 135157.
Broadie, M. and Glasserman, P. (1997) Pricing American-style securities
using simulation. Journal of Economic Dynamics and Control, 21,
13231352.
Cox, J.C., Ross, S.A. and Rubinstein, M. (1979) Option pricing: a simplied approach. Journal of Financial Economics, 7, 229263.
Dasu, S. and Li, L. (1997) Optimal operating policies in the presence of
exchange rate variability. Management Science, 43(5), 705722.
Hodder, J.E. and Triantis, A.J. (1993) Valuing exibility: an impulse control framework. Annals of Operations Research, 45, 109130.
Huchzermeier, A. and Cohen, M.A. (1996) Valuing operational exibility
under exchange rate risk. Operations Research, 44(1), 100113.
Kamrad, B. and Ritchken, P. (1991) Multinomial approximating models for options with k state variables. Management Science, 37(12),
16401653.
Kogut, B. and Kulatilaka, N. (1994) Operating exibility, global manufacturing, and the option value of a multinational network. Management Science, 40(1), 123139.
Majd, S. and Pindyck, R.S. (1987) Time to build, option value, and investment decisions. Journal of Financial Economics, 18, 727.
Marbach, P. and Tsitsiklis, J.N. (2001) Simulation-based optimization of
Markov reward processes. IEEE Transactions on Automatic Control,
46(2), 191209.
McDonald, R. and Siegel, D. (1985) Investment and the valuation of
rms when there is an option to shut down. International Economic
Review, 26(2), 331349.

956
Nembhard, H.B., Shi, L. and Aktan, M. (2002) A real options design for
quality control charts. The Engineering Economist, 47(1), 2859.
Nembhard, H.B., Shi, L. and Aktan, M. (2003) A real options design for
product outsourcing. The Engineering Economist, 48(3), 199217.
Nembhard, H.B., Shi, L. and Park, C.S. (2000) Real option models for
managing manufacturing system changes in the new economy. The
Engineering Economist, 45(3), 232258.
Smith, J.E. and McCardle, K.F. (1998) Valuing oil properties: integrating option pricing and decision analysis approaches. Operations Research, 46(2), 198217.
Trigeorgis, L. (1999) Real Options, MIT Press, Cambridge, MA.
Tsitsiklis, J.N. and Van Roy, B. (1999) Optimal stopping of Markov
processes: Hilbert space theory, approximation algorithms, and an
application to pricing nancial derivatives. IEEE Transactions on
Automatic Control, 44(10), 18401851.
Van Roy, B. and Tsitsiklis, J.N. (2001) Regression methods for pricing
complex American-style options. IEEE Transactions on Neural Networks, 12(4), 694703.

Biographies
Harriet Black Nembhard is an Associate Professor and Bashore Career
Professor in the Harold and Inge Marcus Department of Industrial and
Manufacturing Engineering at The Pennsylvania State University. She is

Nembhard et al.
also the Principal Investigator of the Laboratory for Quality Engineering
and System Transitions (QUEST). Her research interests are in the areas
of applied statistics and nancial engineering with applications to manufacturing, health care, and global sustainability. She received her Ph.D. in
Industrial and Operations Engineering from the University of Michigan
in 1994. She is a member of IIE, INFORMS, ASQ, and a past president
of the Engineering Economy Division of IIE.
Leyuan Shi is a Professor in the Department of Industrial Engineering at
the University of Wisconsin-Madison. She received her Ph.D. in Applied
Mathematics from Harvard University in 1992. Her research interests
include simulation modeling and large-scale optimization with applications to operation planning and scheduling, supply chain management,
transportation, and health care systems. She is serving on the Editorial
Boards of the Journal of Discrete Event Dynamic Systems, INFORMS
Journal on Computing, and Journal of Methodology and Computing in
Applied Probability. She also chairs the Outstanding Publication Award
Committee for the INFORMS College on Simulation.
Mehmet Aktan is an Assistant Professor in the Department of Industrial
University, Turkey. He received his Ph.D. from
Engineering at Ataturk
The University of Wisconsin-Madison, his M.S. from The University of
Iowa, and his B.S. from Bogazici University. He is the co-author of several
research publications in the nancial engineering area.