International Journal of Production Research, 2014

Vol. 52, No. 17, 4950–4966, http://dx.doi.org/10.1080/00207543.2014.885663

Managing production level in new product diffusion: an agent-based simulation approach

A. Negahbana∗ , L. Yilmazb and T. Nallb
a Department of Industrial and Systems Engineering, Auburn University, Auburn, AL, USA; b Department of Computer Science and
Software Engineering, Auburn University, Auburn, AL, USA
(Received 21 January 2013; accepted 15 January 2014)

Managing production level after the launch of a new product is a challenging problem and is critical to the overall profit
of manufacturing firms. The problem involves concepts from different fields including production planning, manufacturing
flexibility, forecasting, and marketing. In this paper, the gaps in the existing literature are first illustrated. With the goal of
addressing the identified research gaps, agent-based modeling and simulation is employed to analyze the performance of
different production planning strategies under various levels of volume flexibility and consumer social network structures.
The key distinguishing feature of the developed model is the capability of the manufacturing firm to adjust its production
level by forecasting the future demand. The analysis of the simulation outputs yields substantial results by challenging
some intuitive and traditional understandings of manufacturing systems. The paper also provides a discussion on managerial
implications of the results in order to provide the managers with guidelines on the implementation of the model in real-world
diffusion environments.
Keywords: production planning; new product diffusion; agent-based modeling and simulation; manufacturing flexibility;
demand forecasting

1. Introduction
During the past few decades, with the increased rate of introduction of new products and product customization, manufacturing
firms have witnessed an ever-increasing competition in the market place due to shorter product life cycles and higher levels
of demand uncertainty. The significance of new products is indicated by the fact that on average roughly one-third of overall
sales and profit of firms comes from new products (Griffin 1997; Hauser, Tellis, and Griffin 2006). Furthermore, due to the
lack of historical sales data, forecasting the demand for a new product during the course of its diffusion becomes very difficult
which in turn makes management of the production level even more challenging. Consequently, manufacturing firms have
realized the importance of volume flexibility and demand projection as the two key elements to cope with this dynamic
environment.
Efficient production planning is vital to manufacturing systems in today’s highly competitive market place. Production
planning involves decision making regarding the timing and magnitude of expansion and reduction in production level relative
to changes in the demand. As a result, it is crucial for manufacturing systems to consider both marketing and production
aspects in order to obtain and maintain a reasonable market share and service level in terms of delivery speed and demand
satisfaction. However, the literature on new product diffusion primarily focuses on developing diffusion models to enhance
demand forecasting by ignoring production constraints and assuming unlimited supply. As a result, the current literature
leaves the following important questions unanswered:
• What is the best strategy to manage production level after the launch of a new product?
• What level of volume flexibility is desirable to maximize the profit from the new product? What level of volume
flexibility minimizes lost sales?
• Is there any correlation between lost sales and profit? Is the traditional understanding that higher flexibility improves
both of these objectives true?
• How sensitive is the profit or lost sales to the individuals’ innovativeness or social network structure in the presence
of supply constraints?
This study aims to fill the gaps in the existing literature by answering the above questions. More specifically, using
Net Present Value (NPV) of profit and lost sales as the two performance measures, we evaluate the performance of three

∗ Corresponding author. Email: [email protected]

© 2014 Taylor & Francis

 International Journal of Production Research 4951

production level management strategies under different volume flexibility levels, social network structures, and diffusion-
related characteristics of the population of potential customers. We employ Agent-Based Modeling and Simulation (ABMS)
technique to develop a model of a virtual market consisting of potential consumers and a manufacturing firm that is modeled
as an adaptive autonomous agent capable of adjusting its production level based on the observed diffusion process. In
other words, in the current study, the manufacturer forecasts the future demand based on its forecasting method and makes
production expansion or reduction decisions based on its production management strategy and current inventory levels. To
the best of our knowledge and based on the results of a recent comprehensive survey on ABMS applications in marketing
(Negahban and Yilmaz 2013), this is the first ABMS study that explicitly incorporates the manufacturing aspect of the market
place by relating new product diffusion to production management, demand forecasting, and volume flexibility.
The remainder of the paper is organized as follows: Section 2 provides a review on new product diffusion literature.
Section 3 presents the conceptual model and its assumptions as well as technical details regarding the development of
the agent-based model. The experimental design and simulation results are discussed in Section 4. A brief discussion on
managerial implications of the findings is presented in Section 5. Finally, conclusions and future research opportunities are
provided in Section 6.

2. Literature review
The importance of new product diffusion has given rise to a huge body of research with early studies going back to the 1960s
(Fourt and Woodlock 1960). In this section, we first identify the shortcomings of the existing literature and then discuss some
recent endeavors to fill these gaps.
The general topic of diffusion has been addressed in several books (Coleman, Katz, and Menzel 1966; Brown 1981;
Rogers 2003; Tidd, Bessant, and Pavitt 2005). In general, two main classes of studies can be identified in the diffusion
literature: (1) analytical models; and (2) simulation studies. Perhaps the most widely used analytical model in the industry
is the Bass model (Bass 1969). Among other commonly used diffusion models are threshold models (Granovetter 1978)
and cascade models (Goldenberg, Libai, and Muller 2001). During the past two decades, simulation techniques, particularly
ABMS, have been extensively applied to study and discover those aspects of new product diffusion that analytical models
are unable to capture. Numerous ABMS studies have been published and analyzed in several review papers (Bonabeau 2002;
Zenobia, Weber, and Daim 2009; Kiesling et al. 2011). However, the majority of the models in both streams ignore production
constraints by assuming unlimited supply (Negahban and Yilmaz 2013).
It was not until recently that a number of diffusion studies have considered both marketing and production facets of the
problem. In the context of analytical models, Kumar and Swaminathan (2003) generalize the Bass model by considering
a firm with a fixed production level and use the proposed model to find production and sales plans that maximize profit
during the lifetime of the product. With the same goal, Xiaoming et al. (2011) present a supply-restricted diffusion model
that accounts for negative WOM from dissatisfied adopters. Cantamessa and Valentini (2000) study the trade-offs among
increasing production level, delaying launch, increasing backlog or accepting a significant amount of lost sales. They develop
a mixed-integer linear programming model and use real data as different demand scenarios in order to find optimal policies
concerning the above factors. Amini and Li (2011) develop a mathematical optimization model to determine the configuration
of the supply chain for new products by optimal selection of production and sales plans as well as timing of the launch of the
new product. An extension to this model is also proposed to accommodate for multiple suppliers and vendors (Li and Amini
2012).
A few ABMS studies also have considered the production side of the diffusion process. Amini et al. (2012) use agent-based
simulation (ABS) to evaluate the performance of myopic and build-up production-sales policies. They study the impact of
negative WOM and length of the build-up period on the NPV of profit. However, in their study, they only implicitly address
supply constraints by assuming a fixed production level. Ma and Nakamori (2005) develop an ABS model to explore the
process of technological innovation as an evolutionary process. The actors in their model include producers and consumers,
where producers are able to adjust the design parameters and make new products that will better meet consumers’ preferences
while there is no restriction on their production level. In another study, Garcia (2005) explores the application of ABMS in
innovation and new product development research though case studies of several electronics manufacturers in the US. In
this model, the only decision that manufacturer agents make is to determine the percentage of resources to be allocated to
research and development while quarterly profitability and market share are used as the performance measures; however,
the demand is assumed to be constant. For a list of related studies on the use of multi-agent systems for investigating supply
network in a non-diffusion context, see the papers by Nair and Vidal (2011) and Lee and Kim (2008).
Although the discussed papers, regardless of their limitations, offer valuable contributions to the field, the presented
literature review demonstrates the existing research gaps in terms of assumptions regarding the supply side of the diffusion
process. This paper contributes to the literature in three significant ways: (1) it provides an evaluation of different strategies
4952 A. Negahban et al.

to manage production level during the diffusion of a new product under various levels of volume flexibility, consumer social
network structures, and innovativeness of the population; (2) it presents an agent-based model in which the manufacturer
agent is capable of forecasting the future demand and adjusting its production level accordingly, i.e. neither demand nor
supply is deterministic or fixed; and (3) it yields substantial results by verifying some of the findings from previous studies
while challenging several intuitive and traditional understandings regarding volume flexibility and the correlation between
profit and lost sales in a diffusion context.

3. Agent-based model development

This section presents the general architecture of the conceptual model along with the ABS algorithm as well as detailed
descriptions of different model components, namely consumer decision making process, social system structure, production
management strategies, and the demand forecasting method.

3.1 Conceptual model

The market place under consideration consists of socially networked consumers that interact with each other through positive
and negative word-of-mouth (WOM) and make decisions regarding the adoption of a new generic product. There is a single
manufacturing firm that produces the product based on a make-to-stock approach. The general structure of the conceptual
model is provided in Figure 1. The resulting market place resembles a complex adaptive system which makes ABMS an
ideal tool to use since it provides a natural description of the social system by modeling the behavioral entities (i.e. the
interacting individuals) and can incorporate population heterogeneity, social network structure, and other micro-level factors
that contribute to the resulting complex non-linear emergent phenomena. The assumptions of the model are as follows:
• Consumers are engaged in a one-time purchase decision.
• Individual’s social network is fixed, i.e. the consumer maintains her relationship with other individuals in her personal
network.
• The manufacturer can change its production level only in discrete time steps. In other words, once the new production
level is set, it cannot be changed until the beginning of the next planning horizon.
• Machine breakdown or other manufacturing disturbances are ignored. Therefore, the yield in each time step will be
equal to the production level at that time step.
• There is no lead time for the supply. Therefore, given the availability of supply, a new adopter will receive the product
instantaneously.

3.2 ABS algorithm

In this model, consumers are divided into three main groups: undecided consumers, adopters, and rejecters. Rejecters are
those consumers that have decided not to adopt the product while undecided consumers are those that have not made a decision
on whether to adopt or reject the new product. There are two major subclasses of adopter consumers namely adopters with
met demand and those adopters whose demand has not been met due to supply shortage. Those adopters that have received
the product, i.e. adopters with met demand, are further categorized into satisfied and dissatisfied subclasses. At each period,
adopters with unmet demand, make a decision on whether they are willing to wait for another period to receive the product.
If they decide to wait for supply, they will be categorized as waiting, otherwise they will be considered as lost customers.
The adoption status of consumers discussed above is adopted from a previous study by Goldenberg et al. (2007) and is
summarized in Figure 2.
The logic of the simulation process can be summarized as follows. The simulation starts by creating a manufacturer agent
with an initial production level and a population of potential customers. The consumers are then connected to each other
based on the choice of the social network. At the beginning of each time period, undecided consumers are influenced by their
neighbors, with whom they have direct social ties, through positive and negative WOM. While rejecters, dissatisfied, and
lost consumers are engaged in negative WOM, satisfied consumers communicate positive WOM. Undecided and waiting
consumers engage in neither positive or negative WOM. Based on the level of positive and negative WOM received as well
as its innovativeness, the undecided consumer then makes a decision as to adopt, reject, or remain undecided.
Once an undecided consumer adopts the product, it will receive the product if there is available supply. The amount of
supply at the beginning of each period is equal to the sum of the current production level and the inventory carried from the
previous period. The supply will be used to satisfy the demand of waiting customers as well as the demand from new adopters.
The manufacturing system has a fixed planning horizon, i.e. discrete time steps at which the production level can be changed.
 International Journal of Production Research 4953

Figure 1. The general architecture of the proposed agent-based model.

Figure 2. Pools of consumers in the virtual market.

At the end of each planning horizon, based on the previous demand levels, the manufacturer forecasts the future demand using
a forecasting method and then sets the production level for the next planning horizon according to its production management
strategy. The simulation stops when all of the individuals in the market have made their adoption/rejection decision.
In this paper, the NPV of profit and the total number of lost consumers are used as the two performance measures. In each
time period, the profit is calculated by subtracting production and inventory costs from the sales in that period. At the end of
the simulation run, the NPV of these profit terms is calculated according to the manufacturer’s Minimum Attractive Rate of
Return (MARR). The total number of lost customers is also collected at the end of each simulation run. Figure 3 provides the
Unified Modeling Language (UML) class diagram and Figure 4 presents the UML sequence diagram of the proposed model
as a formal description of the simulation logic.

3.3 Consumer decision making

Consumer decision making involves concepts from different fields including behavioral and social sciences, psychology,
economics, and marketing. In the context of ABS of virtual markets, several approaches have been proposed to model
consumer decision making process which generally model consumers as autonomous entities that make purchasing decisions
based on static or dynamic rules. For a comprehensive list of these approaches, see the review paper by Negahban and Yilmaz
(2013). The decision making approach along with the state transition equations presented in this section are adopted from
4954 A. Negahban et al.

Figure 3. UML class diagram of the proposed simulation model.

Figure 4. UML sequence diagram of the simulation algorithm.

 International Journal of Production Research 4955

Figure 5. Positive and negative WOM contributors.

the paper by Goldenberg et al. (2007). The approach takes into account both external (e.g. promotions and mass media
advertising) and internal (e.g. positive and negative WOM) influences. The underlying assumptions of this approach are also
consistent with the Bass model of innovation diffusion (Bass 1969) which is widely accepted as a fundamental model of new
product diffusion. This section presents the equations used to model consumer decision making.
In order to calculate the state transition probabilities, we will first determine the probability for consumer i being influenced
by either external advertisement or internal positive and negative WOM. The magnitude of the effect of external marketing
efforts is captured by the coefficient of innovation, p, and the strength of the influence of WOM is captured by the coefficient
of imitation, q. Based on the existing empirical and theoretical studies in the literature, we vary the aggregate parameter p
from 0.01 to 0.05 in 0.01 increments and q from 0.1 to 0.5 in 0.1 increments (Sultan, Farley, and Lehmann 1990; Jiang, Bass,
and Bass 2006). While the value of p will be the same for both individual and aggregate levels, in order to transform q into
individual-level parameter qi , it is divided by the number of neighbors in the individual’s social network (Goldenberg, Libai,
and Muller 2002). Given the above, at any period t N ow, the probability of individual i being influenced by positive WOM
+ −
or advertising (Pi,t N ow ) and negative WOM (Pi,t N ow ) is calculated by:

+
pi,t N ow = 1 − (1 − p) (1 − q j ), where j ∈ L(i)+ , (1)
j

−
pi,t N ow = 1 − (1 − M ∗ q j ), where j ∈ L(i)− . (2)
j

In (1), (1 − p) denotes the probability that consumer i is not influenced by positive effect of advertisement. L(i)+ is
the set individuals who are neighbors to consumer i and communicate positive WOM. The probability of consumer i not
being influenced by a positive WOM contributor j in her social network is (1 − q j ). Therefore, subtracting the product
of these two terms from 1 gives the probability of receiving positive influence regarding the new product. In (2), the term
(1 − M ∗ q j ) denotes the probability that consumer i does not receive any negative WOM in the current period from her
neighbor j, where M indicates that the impact of negative WOM is M times greater than positive WOM. In this paper, M
is chosen to be 2 based on similar studies (Amini et al. 2012). Finally, L(i)− is the set of individuals who are neighbors to
consumer i and communicate negative WOM. As discussed in Section 3.2 and illustrated in Figure 5, rejecters, dissatisfied,
and lost consumers communicate negative WOM while satisfied consumers communicate positive WOM, and undecided
and waiting consumers contribute to neither positive or negative WOM.
We now turn to the calculation of state transition probabilities. At any given time, the probability of consumer i receiving
only positive WOM is (1 − pi− ) pi+ , similarly for only negative WOM is (1 − pi+ ) pi− , and the probability of being influenced
by both positive and negative WOM is given by pi+ pi− . A consumer i who receives both positive and negative WOM adopts
the product with a probability of αi = pi+ /( pi+ + pi− ). The equation for state transition probabilities are therefore given by:
adopt − + + −
pi,t N ow = (1 − pi,t N ow ) pi,t N ow + αi,t N ow pi,t N ow pi,t N ow , (3)
reject + − + −
pi,t N ow = (1 − pi,t N ow ) pi,t N ow + (1 − αi,t N ow ) pi,t N ow pi,t N ow , (4)
+ −
undecided
pi,t N ow = (1 − pi,t N ow )(1 − pi,t N ow ). (5)

Equations (3)–(5) provide the probabilities of adoption, rejection, and remaining undecided, respectively, and would
add up to 1. We define the percentage of dissatisfaction, d, to be the probability of becoming dissatisfied with the product.
adopt
Therefore, an adopter becomes a satisfied customer with a probability of (1 − d) pi,t N ow and becomes a dissatisfied customer
adopt
with a probability of d( pi,t N ow ). Here, based on the literature, the value of d is chosen to be 5% (Goldenberg et al. 2007;
Amini et al. 2012).
4956 A. Negahban et al.

Figure 6. Social network structures.

Finally, if there is no supply available, adopters will have to make a decision as to wait for another period or to completely
withdraw, i.e. become a lost customer. The percentage of backlogged demand is defined as the probability of waiting in case
of supply shortage. The waiting probability is chosen to be 0.5 based on the literature (Amini et al. 2012).

3.4 Consumer social network structure

The structure of the social system has been found as an important element of agent-based models of markets that can have
a significant impact on the resulting market dynamics. Among different types of social system structures, social networks
are the most commonly used method by marketing researchers to model the social world (Negahban and Yilmaz 2013). In
this paper, we consider two theoretical network structures, namely random and regular lattice, as well as two more realistic
network types, namely small-world and scale-free. The structure of these networks are compared in Figure 6 and a brief
description of each network type and the corresponding parameter specifications are provided below.
• Random Network: In a random network, an edge exists between any two vertices with a probability of β. The
advantage of this type of network is that it is easy to implement, which has made it popular among marketing
researchers. Goldenberg et al. (2007) conduct an empirical survey to estimate the number of ties and find that it
varies from 3 to 28 and 1 to 40 for strong and weak ties, respectively. Based on the results from their study, the value
of β in this paper is chosen such that the average number of ties in the random network is 26.
• Regular Lattice: A regular lattice is a network where every node in the network is connected to a particular number of
neighbors. Based on the literature (Stonedahl, Rand, and Wilensky 2010) and in order to provide a common ground
to compare the impact of different network types, the number of neighbors for the regular lattice network is set to 26.
• Small-World Network: A small-world network, as introduced by Watts and Strogatz (1998), is generated by starting
with a regular lattice and then randomly rewiring some of the edges. The small-world network presented in this paper
is generated by a degree of 26 with a rewiring probability of 0.01 (see Stonedahl, Rand, and Wilensky 2010, who
have used the same values for these parameters).
• Scale-Free Network: Unlike other network types, the scale-free network exhibits the preferential connectivity
observed in most real networks (Barabási and Albert 1999). In these networks, the probability for each node of
having n neighbors decays according to a power law with exponent γ . Therefore, few agents have many neighbors
while most of the agents only have a few. To implement the scale-free network, we start with N vertices and
sequentially add edges with a probability relative to the number of existing edges connected to each node. In this
paper, we set N = 10 and γ = 3 according to the guidelines provided by Barabási and Albert (1999).
As for the population size, several researchers suggest that 1000 is sufficiently large to capture the underlying dynamics
of the market and provide reliable results (Cowan and Jonard 2004; Alkemade and Castaldi 2005). Experimental design
studies have also been conducted on larger networks which report that there is no significant difference in the simulation
results obtained from markets of size 3000 and above (Goldenberg et al. 2007; Amini et al. 2012). In the current paper, in
order to guarantee statistically valid results, we model a market consisting of 3600 socially networked consumers.

3.5 Production level management strategies

A strategic decision in a diffusion setting is the timing of changes in the production level, that is, whether it should change
prior to expected changes in demand or after the corresponding level of demand has been realized. In this study, it is assumed
 International Journal of Production Research 4957

Figure 7. Lagging, leading, and tracking demand strategies.

that due to manufacturing constraints, such as procurement of raw material, the manufacturing firm can only change its
production level in discrete time steps. We call these time steps as the planning horizon. The length of the planning horizon
is directly related to the level of volume flexibility of the manufacturing system. As the flexibility of the system increases,
planning horizon can be shorter resulting in a more responsive system to changes in demand levels. In order to be able
to analyze the impact of planning horizon length on the performance of the manufacturing system, we conduct simulation
experiments with fixed planning horizons in the range of 3 to 10 time periods.
Besides timing, the magnitude of changes in production level is the other strategic decision. Olhager, Rudberg, and
Wikner (2001) identify three main strategies used in the literature: (1) lagging demand; (2) leading demand; and, (3) tracking
demand. In this paper, we consider these three strategies and study their performance in a new product diffusion context. A
brief description of each approach and its implementation in the simulation logic is provided as follows:
• Lag Strategy: The general rule for this strategy dictates that production level should not exceed demand as much
as possible. However the strategy could be challenging when demand is declining since the decision to reduce
production level is made when demand is still high (Olhager, Rudberg, and Wikner 2001). In order to implement
this strategy in our simulation model, at each decision point, the production level for the next planning horizon is
set to the most recent demand that has been realized. Therefore, this strategy does not require demand forecasting
and thus is easy to implement. In this way, for an increasing demand, it is always guaranteed that production will
never exceed demand. However, when demand is declining, the same problem discussed above occurs. Figure 7(a)
illustrates how this strategy is applied to respond to changes in demand.
• Lead Strategy: The lead strategy is based on the objective of maintaining an excess production level as a cushion
to respond to changes in demand. The general guideline to implement this strategy is to set the production level in
anticipation of demand. Therefore, this strategy involves forecasting the future demand. However, when demand has
a decreasing pattern, the strategy fails to meet its objectives. Figure 7(b) illustrates how the concept of lead strategy
has been implemented in this study. At each decision point, the production level for the next planning horizon is set
to the forecast of the demand in the last period of the next planning horizon.
• Track/Chase Strategy: This strategy is developed to combine positive aspects of the two pure strategies described
above. The tracking demand strategy aims at finding an efficient way to minimize the deviations between production
and demand by tracking the demand as close as possible which requires forecasting future demand levels. To
implement this strategy in the simulation algorithm, the production level for the next planning horizon, n, is chosen
to be the average of the most recent observed demand, d, and the forecast of future demand in the last period of the
next planning horizon, f . This strategy is shown in Figure 7(c).
At the end of each planning horizon, when the manufacturer agent is activated, it forecasts the demand based on the
observed demand levels in previous periods using a given forecasting method. Then, the new production level for the next
planning horizon is determined based on the production management strategy. Finally, the last 16% of the consumers that
adopt the product, also known as laggerds, are very slow in adopting and this may result in unnecessary manufacturing and
inventory costs which in turn can affect our evaluation of different production strategies. Therefore, in order to provide robust
results, we solve this problem by applying the same approach as proposed by Amini et al. (2012). In order to implement
4958 A. Negahban et al.

their approach in our model, the manufacturer switches to lagging demand when 84% of potential customers have made their
adoption decision.

3.6 Demand forecasting method

In this model, we study the performance of the Bass model as the forecasting method used by the manufacturing firm. The
Bass model was first introduced by Bass (1969) and became one of the most influential models in management science. There
have been numerous applications of the Bass model in the industry and the majority of the diffusion models are essentially
rooted in this method (Bass 2004). The model is developed for products involving one-time purchase decisions where the
sales in each time period t, S(t), is a function of the coefficient of innovation, p, coefficient of imitation, q, and the total
number of purchases in the (0, t) interval, Y (t). Since we are dealing with consumer durable products, i.e. products with
one-time purchase, Y (t) will be equal to the total number of adopters up to time t. We have

S(t) = pm + (q − p)Y (t) − q/m[Y (t)]2 , (6)

where m is a constant. The reader is referred to the original paper for more details on the theory behind this model. To
implement the Bass model, at the end of each planning horizon, the demand history from the last three periods is first used
to estimate the parameters of the model by solving a system of equations. Equation (6) is then used in a recursive process
to forecast the demand of the periods in the next planning horizon. The forecast is adjusted by incorporating the number of
backlogged or waiting customers with unmet demand from the previous periods. The demand forecast is then fed into the
production strategy to make decision on the new production level. This concludes the description of different components
of the simulation model.

4. Model implementation, experimental design, and results

The selected parameter values and ranges are summarized in Table 1. We define each parameter combination as a scenario.
Therefore, based on the number of levels for the six design factors (production strategy, length of planning horizon, coefficient
of innovation, coefficient of imitation, network type, and discount factor) there are a total of 3 × 8 × 5 × 5 × 4 × 4 = 9600
scenarios. An analysis on the half-width values for the results from several pilot runs indicates that the minimum number
of replications necessary is 30. Therefore, a total of 288,000 replications of different parameter combinations is run. The
simulation model is developed in Repast Symphony which is a Java-based ABS platform (North, Collier, and Vos 2006). All
simulation experiments are run on a standard Dell™ desktop with a 3.00GHz quad-core CPU and 8GB of RAM.

4.1 Model verification and validation

Before conducting the simulation experiments, validation and verification (V&V) are performed according to the guidelines
available in the literature (Balci 1994; Sargent 2005; Rand and Rust 2011) to increase our confidence in the correctness of the
model. Since verification and validation are beyond the scope of this study and for the sake of conciseness, we only provide a
brief description of the V&V steps. The computer model is tested to ensure that it is error-free. Using structured walk-through
and trace techniques, different simulation functions such as time-flow are tested to be correctly implemented. Animation is
used to graphically monitor the behavior of individuals in the population. Degenerate tests are then performed to observe the
behavior of the model in extreme conditions. Operational graphics are also used to dynamically display the output measures
as the model is running to verify the calculations. Finally, sensitivity analysis is performed on a few predictable scenarios
to ensure that output trends match our expectations as discussed below. These runs are based on lagging demand as the
production management strategy, regular lattice network, and planning horizon of 3 time steps, p = 0.03, and q = 0.4 where
all other variables are set to their default values.
• Given all other parameters remain unchanged, it is expected that an increase in the percentage of dissatisfaction, d,
will yield more rejecters due to the increase in negative WOM contributers. We vary the value of d from 0 to 0.25
to observe its effect on the number of rejecters which matches our expectation as shown in Figure 8(a).
• Number of rejecters is also expected to increase if the relative power of negative WOM to positive WOM increases.
This hypothesis is verified by Figure 8(b), where M is varied in the range of 1–4.
• Decreasing the probability of waiting or backlogging is expected to increase lost sales since consumers will be less
likely to wait for future supply. This hypothesis is tested and verified by changing the backlogging probability from
0 to 1 (Figure 8(c)).
Table 1. Parameter choices for the ABS model.

Class Parameter Value/Range Reference

Consumer Coefficient of innovation ( p) 0.01, 0.02, 0.03, 0.04, 0.05 Sultan, Farley, and Lehmann (1990), Jiang, Bass, and Bass (2006)
Coefficient of imitation (q) 0.1, 0.2, 0.3, 0.4, 0.5 Sultan, Farley, and Lehmann (1990), Jiang, Bass, and Bass (2006)
Relative power of negative WOM to positive WOM (M) 2 Amini et al. (2012)
Percentage of dissatisfaction (d) 5% Amini et al. (2012), Goldenberg et al. (2007)
Probability of waiting 0.5 Amini et al. (2012)
Manufacturer Production strategy Lag, Lead, Track Olhager, Rudberg, and Wikner (2001)
Forecasting method Bass Model Bass (1969)
Unit selling price 1.1 Amini et al. (2012)
Unit variable cost 1 Amini et al. (2012)
Unit inventory holding cost 0.01 Amini et al. (2012)
Planning horizon length 3 − 10
Initial production level 40 Amini et al. (2012)
Discount factor (MARR) 0.001, 0.005, 0.01, 0.015 Amini et al. (2012)
Social Network Network type Random, Lattice,
Small-world, Scale-free Negahban and Yilmaz (2013)
Population size 3600 Goldenberg et al. (2007)
International Journal of Production Research

Average number of ties in random network 26 Goldenberg et al. (2007)

Number of neighbors in the regular lattice 26 Goldenberg et al. (2007)
Degree of the small-world network 26 Goldenberg et al. (2007)
Rewiring probability of small-world network 0.01 Stonedahl, Rand, and Wilensky (2010)
Initial number of attachments in scale-free network (N ) 10 Barabási and Albert (1999)
Power law exponent of scale-free network (γ ) 3 Barabási and Albert (1999)
4959
4960 A. Negahban et al.

Figure 8. Results of the validation runs.

• We expect the diffusion to slow down as a result of a higher dissatisfaction probability since it would yield a larger
population of dissatisfied consumers who participate in negative WOM. A slower diffusion process would eliminate
dramatic surges in the demand level, which will reduce lost sales. This effect is tested and simulation results are
inline with our expectation, see Figure 8(d).

4.2 Analysis of simulation results

Once the simulation model is developed, verified and validated, we use it to study the performance of the three production
level strategies under different volume flexibility, social network structures, coefficient of innovation and imitation, and
discount factor values. The performance measures used in this paper are the NPV of profit generated over the diffusion
process and the total number of lost customers. This section presents the results and outlines the important findings, while a
complete discussion on the implications of the simulation results is provided in Section 5.

4.2.1 Production level strategies and volume flexibility

Figure 9 illustrates the average performance of different combinations of production level strategy and volume flexibility
over the four network structures considered in this paper. As shown in the figure, tracking demand strategy performs best in
terms of average NPV virtually regardless of the length of the planning horizon (volume flexibility). This is inline with our
expectation since this strategy is expected to minimize the deviation between production and demand levels while the other
two pure strategies would compromise one for the other. For the case of the leading demand strategy, unlike tracking and
lagging demand strategies, the choice of the planning horizon has a significant impact on the expected NPV. An interesting
observation is that lagging demand strategy performs better than leading demand with respect to lost sales. This can be
attributed to the fact that for the case of decreasing demand (similar to one that happens later in the diffusion process after
the peak), lagging demand provides supply surplus while leading demand can yield to lost sales.
Perhaps the most important finding is related to the intuitive and traditional understanding that high levels of volume
flexibility (i.e. shorter planning horizons) would enable the manufacturing system to adjust more quickly to the changing
demand and thus result in (1) reduced lost sales and (2) increased profit. Our results support the first part pertaining to the
 International Journal of Production Research 4961

Figure 9. Performance of production management strategies and the impact of volume flexibility.

impact of volume flexibility on lost sales. As shown in the figure, under all three strategies, longer planning horizons result in
higher lost sales. However, the results of our simulation study challenge the second part of this traditional view by revealing
that for the case of the leading demand strategy, the expected NPV increases as the planning horizon becomes longer. This
interesting observation can be justified by the fact that the Bass model does not take into account negative WOM which could
result in over-estimating the future demand and consequently, incur unnecessarily high manufacturing and inventory costs
early in the diffusion process. Shorter planning horizons further amplify this problem, as shown in the figure. Therefore,
when using the leading demand strategy along with the Bass model as the forecasting method, longer planning horizons are
preferable.

4.2.2 Population-related factors: coefficients of innovation and imitation

It is well known from the diffusion literature that coefficients of innovation ( p) and imitation (q) have a significant impact
on diffusion speed as well as timing and magnitude of the peak demand. On the other hand, because of the lack of sales
data for a new product, companies do not have an accurate estimate of p and q in most cases. In order to provide a robust
analysis, we investigate the performance of the three production management strategies under different levels of p and
q (Figure 10). Coefficient of innovation increases the level of innovativeness of the population. As the number of early
adopters increases, during the early stages of diffusion, we expect the demand levels to be closer to the forecasts from
the Bass model (remember the Bass model tends to over-estimate the demand since it does not consider negative WOM).
Therefore, as demand forecasts become more accurate, the expected NPV of profit is expected to increase, which also matches
the simulation results. However, as time evolves, the diffusion process is mainly affected by imitators rather than innovators
and thus negative and positive WOM begin to play a crucial role. On the other hand, from Equation (2), it can be derived
that a bigger q, will increase the probability of an imitator being influenced by negative WOM. Consequently, after the early
stages of diffusion, a higher q value reduces the accuracy of predictions of the Bass model more severely resulting in demand
forecasts to be much higher than the actual future demand. Therefore, the expected profit is expected to decrease because of
higher manufacturing and inventory costs while the number of lost customers is expected to decrease due to over-estimating
the future demand.

4.2.3 Consumer social network structure

The results support the findings from previous studies, which suggest that the structure of the social network can have a
significant impact on diffusion dynamics. Figure 11 indicates that the social network structure in which consumer interactions
take place is critical to the firm’s NPV and lost sales after introducing a new product into the market. The results suggest
that under all network structures, the tracking and lagging demand strategies give the best results for the average NPV and
number of lost customers, respectively. Therefore, given a performance measure, the structure of the social network seems
not to change the choice of the production management strategy.
4962 A. Negahban et al.

Figure 10. The impact of coefficients of innovation and imitation on NPV and lost sales.

Figure 11. The impact of consumer social network structure on NPV and lost sales.

4.2.4 NPV vs. lost sales

An important observation from Figures 9–11 is that fewer lost customers does not necessary yield higher profits. This finding
motivated us to further study the relation between these two performance measures. Figure 12 plots the average NPV of the
9600 scenarios vs. their corresponding average number of lost customers. The plot is obtained by ordering all scenarios in the
increasing order of their reported NPV. Although statistical analysis does not indicate that the two performance measures are
 International Journal of Production Research 4963

Figure 12. Average NPV vs. average lost sales.

Figure 13. Average NPV and lost sales under different discount factors.

correlated, an interesting observations can be derived from the graph which indicates some of the scenarios that yield fairly
high average NPV values also have the highest levels of lost sales (points contained in the blue oval). On the other hand,
the minimum lost sales are reported by those scenarios with highly negative NPV values (points contained in the red oval).
Further investigation of this finding reveals that what matters most with respect to the NPV is not simply the number of lost
customers but at what point in time the loss takes place. In other words, our analysis suggests that in order to gain high profit
from the diffusion of the new product, it is vital not to lose too many customers at the early stages of the diffusion process
before the peak demand. This suggests that maintaining reasonably high production levels in the early phases of diffusion
could be beneficial to the manufacturing firm.
This finding also matches the results of a previous study by Kumar and Swaminathan (2003). They propose an extension
to the Bass model to account for supply constrains. The results of their analytical model indicate that it is not optimal
to delay demand fulfillment although this itself could accelerate the diffusion process and thereby result in losing more
customers. However, we challenge the results from previous studies that attribute reduced profit directly to the level of lost
sales (Cantamessa and Valentini 2000). In fact, we show that, from the financial point of view, losing customers only in the
early phases after the product launch could have a significant impact on the firm’s profit. In other words, reduced profits are
not due to lost sales in general but due to lost sales that occur early in the diffusion process.

4.2.5 Discount factor

Figure 13 illustrates the results for different discount factor values. The results indicate that for all discount factor values,
the tracking demand strategy gives the best expected NPV of profit. Furthermore, as expected, discount factor does not have
an impact on the number of lost customers since it only affects NPV calculations.
4964 A. Negahban et al.

5. Discussion
In this section we discuss some managerial implications concerning the implementation of the proposed model. We strongly
believe managers of manufacturing companies can benefit from these recommendations by better understanding the under-
lying dynamics and interdependencies of new product diffusion and production management strategies.
• The current study considers NPV of profit and lost sales as the two performance measures and illustrate that the
two criteria could be in conflict. In general, when monetary and quantitative criteria are to be considered along with
non-monetary criteria, there is a need for analytical methods to select the best solution (Karsak and Ahiska 2005).
Since this is out of the scope of this paper, we do not conduct such analysis. However, before implementing the
results of this study in real-world (i.e. making decision about the best production strategy and volume flexibility),
the managers need to use any of the available decision making tools such as data envelopment analysis (DEA),
multi-criteria decision making (MCDM) methods, goal programming, outranking methods, etc. to be able to select
a strategy that performs well when the two criteria are considered simultaneously. For a review and comparison of a
number of these methods and their applications, see (Mansouri, Husseini, and Newman 2000).
• Highly flexible production systems are extremely capital intensive. Moreover, scheduling and control of manufac-
turing systems become more challenging as the flexibility increases. Nevertheless, flexibility brings a wide range
of advantages to manufacturing systems that were not addressed in our analysis. Here, we name a few. Flexible
manufacturing systems are less vulnerable to interruptions in production including machine breakdowns, blocking
and starving, tool changeovers and setups, preventive maintenance, etc. Labor costs decrease with higher levels of
automation in flexible systems. Flexibility allows for production of multiple product types, which eliminates the need
for having multiple production lines and thus help manufacturing firms save a considerable amount of money on land
and overhead costs. Moreover, such systems will not require major modifications as the new product evolves during
its life cycle. As a result, far more investigation is necessary to thoroughly analyze these trade-offs in order for the
managers to be able to make the right decision on the required level of flexibility before acquiring the production
line for a new product.
• Our results indicate that the structure of the social network can play a key role in the NPV of profit obtained from
the introduction of a new product. Although the diffusion under different traditional network structures have been
the focus of a vast body of the marketing literature, we suggest the use of more realistic network structures. Today,
there is a massive amount of information available about the structure of individuals’ social ties through various
social networking websites such as Facebook, Twitter, LinkedIn, etc., providing managers with the opportunity for
transition from theoretical structures to more realistic social networks. We recommend the use of empirical network
structures to increase the robustness and validity of their analysis.
• Although the consumer decision making model used in this study has strong theoretical background and can provide
useful results and helpful insights, practitioners may find its assumptions unrealistically simplified. Similarly,
managers might be skeptical about the validity of other commonly used behavioral models. Nowadays, with the
advancement of surveying technology and online marketing, companies have access to a wide spectrum of data
about current or potential customers regarding their age, gender, income, marital and housing status, and other
personal or socio-demographic attributes. Here, we do not dispute the contributions of the currently used methods to
the literature, but we would argue that the use of such information along with more sophisticated behavioral decision
making models that are based on evolving psychological theories of behavior would allow companies to fully utilize
the potential of the microdata available to them.
• In this paper, we consider the production process in an aggregate level. In other words, we do not include the details of
the production system into the model. We have also assumed that the production of the new product is not interrupted
by any of the interruptions mentioned above; however, in reality, these factors would impact the production yield.
Therefore, if the managers believe that these uncertainties have a significant impact on their production level such
that the results would become invalid if ignored, they can expand the proposed architecture by adding more details
about the manufacturing process into the model to account for these uncertainties. Moreover, the model could be
expanded to account for multiple manufacturers in network-centric manufacturing environments (Oh and Shin 2012).

6. Conclusions
In this paper, an analysis of different strategies for managing production level in new product diffusion is presented. NPV of
profit and lost sales are used as the two performance measures. An agent-based model is developed where the diffusion takes
place among socially networked consumers and the manufacturing system adjusts its production level based on its volume
flexibility and forecasts of the future demand levels.
 International Journal of Production Research 4965

Some of the results of our simulation study support previous findings in the literature. For instance, we show that the
structure of the consumer social network plays a significant role in the emerging diffusion dynamics. In addition, we show
that higher levels of volume flexibility reduce the number of lost customers since it enables the manufacturing system to
adjust its production level more quickly to the variations in demand. As expected and supported by our results, tracking
demand strategy has a better performance than the two pure strategies considered in this study with respect to NPV of profit.
We also verify the results of a few existing papers on supply-constrained innovation diffusion by showing that delaying the
sales is not beneficial to the manufacturing firm.
The results also challenge several conventional beliefs. Simulation results suggest that for the leading demand strategy,
the best NPV values are realized at lower flexibility levels, in contrast to the traditional understanding that higher flexibility is
always desirable. We characterize the relation between the two performance criteria and highlight the interesting observation
that some of the scenarios that report very high NPV values also have the highest lost sales, which challenges the intuitive
understanding that high profits correspond to fewer lost customers. Our analysis reveals that reduced profits are not directly
related to the number of lost customers in general but to lost sales that occur in the early stages of the product launch.
The proposed model involves several assumptions and certainly does not capture all of the dimensions of new product
diffusion. The model could be improved by exploring various initial conditions and relaxing some of the assumptions. The
sensitivity of the results to the percentage of dissatisfaction, probability of backlogging, and other parameters could also be
investigated. The dynamics of the problem could be studied under other types of theoretical or empirical network structures,
forecasting methods, production strategies, and pricing approaches. Future research could also enhance the consumer decision
making process by incorporating more realistic and complex behavioral rules. Real-world market consists of multiple
manufacturing firms that compete with each other to achieve higher market share. Therefore, another interesting topic
for future research would be the development of a model consisting of multiple manufacturer agents. The authors strongly
believe that addressing the above issues would provide researchers and practitioners with valuable insights and a deeper
understanding of the underlying dynamics of new product diffusion.

References

Alkemade, F., and C. Castaldi. 2005. “Strategies for the Diffusion of Innovations on Social Networks.” Computational Economics 25
(1–2): 3–23.
Amini, M., and H. Li. 2011. “Supply Chain Configuration for Diffusion of New Products: An Integrated Optimization Approach.” Omega
39 (3): 313–322.
Amini, M., T. Wakolbinger, M. Racer, and M. G. Nejad. 2012. “Alternative Supply Chain Production-sales Policies for New Product
Diffusion: An Agent-based Modeling and Simulation Approach.” European Journal of Operational Research 216 (2): 301–311.
Balci, O. 1994. “Validation, Verification, and Testing Techniques Throughout the Life Cycle of a Simulation Study.” Annals of Operations
Research 53 (1): 121–173.
Barabási, A. L., and R. Albert. 1999. “Emergence of Scaling in Random Networks.” Science 286: 509–512.
Bass, F. M. 1969. “A New Product Growth for Model Consumer Durables.” Management Science 15 (5): 215–227.
Bass, F. M. 2004. “Comments on “A New Product Growth for Model Consumer Durables.” Management Science 50 (12): 1833–1840.
Bonabeau, E. 2002. “Agent-based Modeling: Methods and Techniques for Simulating Human Systems.” Proceedings of the National
Academy of Sciences of the United States of America 99: 7280–7287.
Brown, L. A. 1981. Innovation Diffusion: A New Perspective. New York: Methuen.
Cantamessa, M., and C. Valentini. 2000. “Planning and Managing Manufacturing Capacity When Demand is Subject to Diffusion Effects.”
International Journal of Production Economics 66 (3): 227–240.
Coleman, J., E. Katz, and H. Menzel. 1966. Medical Innovation: A Diffusion Study. Indianapolis (IN): Bobbs-Merrill.
Cowan, R., and N. Jonard. 2004. “Network Structure and the Diffusion of Knowledge.” Journal of Economic Dynamics and Control 28
(8): 1557–1575.
Fourt, L., and J. Woodlock. 1960. “Early Prediction of Market Success for New Grocery Products.” Journal of Marketing 25 (2): 31–38.
Garcia, R. 2005. “Uses of Agent-based Modeling in Innovation/New Product Development Research.” Journal of Product Innovation
Management 22 (5): 380–398.
Goldenberg, J., B. Libai, S. Moldovan, and E. Muller. 2007. “The NPV of Bad News.” International Journal of Research in Marketing 24
(3): 186–200.
Goldenberg, J., B. Libai, and E. Muller. 2001. “Talk of the Network: A Complex Systems Look at the Underlying Process of Word-of-
mouth.” Marketing Letters 12 (3): 211–223.
Goldenberg, J., B. Libai, and E. Muller. 2002. “Riding the Saddle: How Cross-market Communications can Create a Major Slump in
Sales.” Journal of Marketing 66 (2): 1–16.
Granovetter, M. 1978. “Threshold Models of Collective Behavior.” American Journal of Sociology 83 (6): 1420–1443.
Griffin, A. 1997. “Pdma Research on New Product Development Practices: Updating Trends and Benchmarking Best Practices.” Journal
of Product Innovation Management 14 (6): 429–458.
4966 A. Negahban et al.

Hauser, J., G. J. Tellis, and A. Griffin. 2006. “Research on Innovation: A Review and Agenda for Marketing Science.” Marketing Science
25 (6): 687–717.
Jiang, Z., F. M. Bass, and P. I. Bass. 2006. “Virtual Bass Model and the Left-hand Data-truncation Bias in Diffusion of Innovation Studies.”
International Journal of Research in Marketing 23 (1): 93–106.
Karsak, E. E., and S. S. Ahiska. 2005. “Practical Common Weight Multi-criteria Decision-making Approach with an Improved
Discriminating Power for Technology Selection.” International Journal of Production Research 43 (8): 1537–1554.
Kiesling, E., M. Gunther, C. Stummer, and L. Wakolbinger. 2011. “Agent-based Simulation of Innovation Diffusion: A Review.” Central
European Journal of Operations Research 20 (2): 183–230.
Kumar, S., and J. M. Swaminathan. 2003. “Diffusion of Innovations Under Supply Constraints.” Operations Research 51 (6): 866–879.
Lee, J. H., and C. O. Kim. 2008. “Multi-agent Systems Applications in Manufacturing Systems and Supply Chain Management: A Review
Paper.” International Journal of Production Research 46 (1): 233–265.
Li, H., and M. Amini. 2012. “A Hybrid Optimisation Approach to Configure a Supply Chain for New Product Diffusion: A Case Study of
Multiple-sourcing Strategy.” International Journal of Production Research 50 (11): 3152–3171.
Ma, T., and Y. Nakamori. 2005. “Agent-based Modeling on Technological Innovation as an Evolutionary Process.” European Journal of
Operational Research 166 (3): 741–755.
Mansouri, S. A., S. M. Husseini, and S. Newman. 2000. “A Review of the Modern Approaches to Multi-criteria Cell Design.” International
Journal of Production Research 38 (5): 1201–1218.
Nair, A., and J. M. Vidal. 2011. “Supply Network Topology and Robustness Against Disruptions – An Investigation using Multi-agent
Model.” International Journal of Production Research 49 (5): 1391–1404.
Negahban, A., and L. Yilmaz. 2013. “Agent-based Simulation Applications in Marketing Research: An Integrated Review.” Journal of
Simulation. doi:10.1057/jos.2013.21.
North, M. J., N. T. Collier, and J. R. Vos. 2006. “Experiences Creating Three Implementations of the Repast Agent Modeling Toolkit.”
ACM Transactions on Modeling and Computer Simulation 16 (1): 1–25.
Oh, S. C., and J. Shin. 2012. “A Semantic e-Kanban System for Network-centric Manufacturing: Technology, Scale-free Convergence,
Value and Cost-sharing Considerations.” International Journal of Production Research 50 (19): 5292–5316.
Olhager, J., M. Rudberg, and J. Wikner. 2001. “Long-term Capacity Management: Linking the Perspectives from Manufacturing Strategy
and Sales and Operations Planning.” International Journal of Production Economics 69 (2): 215–225.
Rand, W., and R. T. Rust. 2011. “Agent-based Modeling in Marketing: Guidelines for Rigor.” International Journal of Research in
Marketing 28 (3): 181–193.
Rogers, E. M. 2003. Diffusion of Innovations. 5th ed. New York: Free Press.
Sargent, R. G. 2005. “Verification and Validation of Simulation Models.” In Proceedings of the 37th Winter Simulation Conference, Orlando,
FL, 130–143.
Stonedahl, F., W. Rand, and U. Wilensky. 2010. “Evolving Viral Marketing Strategies.” In Proceedings of the 12th Annual Conference on
Genetic and Evolutionary Computation, Portland, OR, 1195–1202.
Sultan, F., J. U. Farley, and D. R. Lehmann. 1990. “A Meta-analysis of Applications of Diffusion Models.” Journal of Marketing Research
27 (1): 70–77.
Tidd, J., J. Bessant, and K. Pavitt. 2005. Managing Innovation: Integrating Technological, Market and Organizational Change. 3rd ed.
Chichester: Wiley.
Watts, D. J., and S. H. Strogatz. 1998. “Collective Dynamics of ‘Small-world’ Networks.” Nature 393 (6684): 440–442.
Xiaoming, Y., P. Cao, M. Zhang, and K. Liu. 2011. “The Optimal Production and Sales Policy for a New Product with Negative Word-of-
mouth.” Journal of Industrial and Management Optimization 7 (1): 117–137.
Zenobia, B., C. Weber, and T. Daim. 2009. “Artificial Markets: A Review and Assessment of a New Venue for Innovation Research.”
Technovation 29 (5): 338–350.
Copyright of International Journal of Production Research is the property of Taylor & Francis
Ltd and its content may not be copied or emailed to multiple sites or posted to a listserv
without the copyright holder's express written permission. However, users may print,
download, or email articles for individual use.