Chaos Theory and the

Dynamics of Marketing Systems

Brynn Hibbert
University of New South Wales

lan F. Wilkinson
University of Western Sydney, Nepean

Chaos theory is not well understood or appreciated in Chaos is the name given to this unstable yet bounded
marketing, yet it offers a way of improving understanding behavior. It may be defined as a pattern of behavior over
of marketing systems. A brief introduction to chaos theory time that is generated by a deterministic equation but
is given. Its relevance to marketing is then illustrated which is extremely sensitive to the starting conditions such
that no matter how close two starting conditions are they
through a discussion of different types of marketing models
will diverge exponentially over time. Usually we expect
that can result in complex dynamics, including chaos.
that systems that start in very similar conditions will
Next, techniques for detecting the presence of chaos in the behave similarly over time, but this is not the case in chaos.
behavior of real systems are reviewed. Finally, the impli- A consequence of this behavior is the impossibility of
cations of chaos theory for marketing theory and practice long-term prediction, which has important implications for
are discussed. a firm's forecasting and planning behavior.
The existence of complex dynamics in relatively simple
nonlinear systems has been known since the end of the
19th century and the pioneering work of Poincare. It is only
recently, with the advent of modern computer systems, that
Simple nonlinear equations of the type commonly used the full extent and implications of his work have been able
in marketing models can produce unstable yet persistent to be explored.
patterns of behavior over time that appear to be random. The discovery that simple nonlinear difference equa-
These complex dynamics are completely overlooked by tions may lead to complex dynamics has led researchers in
traditional methods of analysis that focus on equilibri- many disciplines to investigate the possibility that basic
um conditions and linear approximations. For example, models of behavior could show such dynamics. This has
Gordon and Greenspan (1988) note that resulted in insights that have important implications for
understanding and controlling systems' behaviors. Exam-
Most systems analysis . . . assumes that a system pies are found in the natural sciences (e.g., Lundquist,
must be stable because only stable behavior persists; Marsh, and Tosi 1988; Prigogine and Stengers 1984;
an unstable system exhibiting equilibrium soon "ex- Campbell and Rose 1983) as well as in economics and
plodes" and therefore, is only of transient interest. finance (e.g., Stutzer 1980; Brock 1986; Brock and Sayers
In chaos analysis unstable but bounded behavior can 1988; Benhabib and Day 1981; Day 1982, 1983a,b;
be demonstrated to persist, which therefore adds a Chiarella 1988; Gordon and Greenspan 1988; Weiss 1992;
new dimension to the analysis of nonlinear systems. Medio and Gallo 1993; Nichols 1993), urban geography
(p. 2) (e.g., Allen and Sanglier 1978, 1979; Beaumont, Clarke,
and Wilson 1981; Clarke and Wilson 1983; Wilson 1981)
and business (e.g., Preismeyer and Batik 1989).
Journal of the Academy of Marketing Science. To date, there has been no work focusing on the role
Volume 22, No. 3, pages 218-233. and importance of chaos theory in marketing, yet several
Copyright 9 1994 by Academy of Marketing Science. important implications for management, research, and the-
 Hibbert and Wilkinson/ CHAOS THEORY 219

ory arise. For managers and policymakers, implications FIGURE 1

arise in terms of the ability to predict and control system Sensitivity to Starting Conditions
behavior. First, chaos theory suggests new ways of under- of Chaotic Time Series
standing and dealing with complex, apparently disorderly
patterns of behavior over time observed in such phenom-
ena as sales, inventories, brand shares, and prices. Tradi-
tional explanations of such phenomena tend to rely on
assumed patterns of external shocks, random noise, or 4-
9| 9 ~4""
inherently stochastic processes to account for the complex
dynamics. Chaos theory offers the opportunity to explain
at least a component of this behavior in terms of determi-
nistic feedback mechanisms reflected in the rules govern-
ing system members' behavior and interactions. This can
lead to new ways of coping with or avoiding these patterns
of behavior, to the extent the rules governing behavior are
I I I I I I I t I I I I I I I I I I I I
amenable to control by firms and/or policymakers. Sec-
lime
ond, chaos theory shows why it may be impossible to make
accurate long-term predictions of system behavior even
though the rules governing that behavior are known.
Implications also arise for research and theory building. AN INTRODUCTION TO CHAOS THEORY
First, research efforts can be better directed toward discov-
ering the mechanisms underlying system behavior. Chaos Equations that show chaos can be very simple. Consider
theory offers a way of distinguishing between purely ran- the following logistic difference equation in which x varies
dom behavior in which prediction is impossible and be- between 0 and 1:
havior for which short-term but not long-term prediction
is possible. This means that situations can be identified in
xt+1 = rxt(1 - xt). [1]
which apparently random behavior may result, at least in
part, from some form of underlying nonlinear feedback The value of the variable of interest (x) in the next time
mechanism. As a result, research efforts can be directed period (xt+ 1) is determined by the value in the current time
toward discovering these mechanisms. period (x~ and a constant r. This is a very general form of
Second, chaos theory offers an alternative explanation the logistic equation to which other forms may be reduced. 1
for the existence of various types of marketing institutions The behavior of x over time is chaotic for values of r
as "disequilibrium mechanisms" designed to buffer or between 3.57 and 4. In Figure 1 we show the behavior of
reduce the effects of complex dynamics. These include x over time for initial values o f x that differ by one percent
inventory-holding intermediaries, financial intermediar- (0.600 and 0.605) and for r = 3.57. No matter how close
ies, insurance agencies, and ordering systems. two initial states are, their trajectories will ultimately be-
Last, chaos theory leads to new ways of explaining and come far apart. This, as we shall see, has important impli-
predicting structural change and evolution in marketing cations for the limits of our ability to predict.
systems. These arise as a consequence of marketing system The parameter r in effect governs the degree of non-
participants modifying the rules governing their behavior linearity of the equation. If the pattern of behavior ofx over
and interaction in response to the emergence of chaotic time is plotted for different values of r, when r = 2 a single
regimes of behavior. equilibrium value o f x emerges (x = 0.5). As r is increased,
The purpose of this article is to provide an introduction overshooting of the equilibrium value occurs, resulting in
to chaos theory and to indicate its potential relevance for oscillations that diminish over time. At r -- 3, the equilib-
the study and practice of marketing. In the next section, we rium becomes unstable and a persistent oscillating pattern
provide a brief overview of some of the important concepts of behavior emerges, with x hopping between two values
of chaos theory. Following this, we show how three types with each iteration. As r is increased further, bifurcations
of marketing models can result in complex dynamics in- occur of the equilibrium values, resulting in successive
cluding chaos. We first examine some existing models of doublings of the oscillations (to 4, 8, 16, etc.). These
product diffusion and market evolution and then develop bifurcations as r increases also occur at more frequent
a simple nonlinear model of brand competition for the intervals (3.45, 3.54, 3.56) until chaos begins at r = 3.57,
purposes of illustrating the ways such models can be as depicted in Figure 1. Beyond this value "there are an
developed and analyzed and how complex dynamics can infinite number of fixed points with different periodicities,
arise under plausible conditions. We then consider how the and an infinite number of different periodic cycles. There
presence of chaos can be detected in the behavior of actual is also an uncountable number of initial points.., that give
systems over time. Finally, we discuss the implications of totally aperiodic (although bounded) trajectories; no mat-
chaos theory for marketing theory and practice and suggest ter how long the time series generated.., is run out" (May
future research directions. 1976, p. 461).
220 JOURNALOF THE ACADEMYOF MARKETINGSCIENCE SUMMER 1994

Feigenbaum (1979, 1983) and others have shown the the product adopt it independently of existing adopters'
universal nature of the period doubling and chaotic sce- influence. The second process is the rate of imitation and
nario. Any one variable equation of the form xt + i = f(x,, r) is reflected in the second part of the equation, qN,(1 -
that has an isolated maximum will have these properties. Nt/m). The rate of imitation refers to the influence of
An isolated maximum means that the function relating the existing adopters on new adopters through word-of-mouth
value o f x in one period to its value in the next has a single and demonstration effects, for example. Generally, such
"humped" shape. As many researchers have pointed out, imitation is assumed to be positive (i.e., q > 1). The
single humped functions of this type are rather common equilibrium situation occurs when N, = m and the number
and have been detected in nearly all branches of the natural of new adopters per period is zero.
and social sciences. Feigenbaum (1979) also shows that, Equation 2 can result in complex dynamics depending
for sets of equations with two or mo'?e variables that on the values o f p and q. This cannot occur given the usual
undergo period doubling, after enough iterations a single assumptions made for this model, which are that m is an
variable will emerge to follow the simple one-variable upper limit that cannot be exceeded and adoption refers
equation pattern. 2 only to first purchases, such that people cannot "unadopt"
once they have adopted the product. As aresult, the number
of adopters per period must be monotonically increasing,
MARKETING MODELS AND CHAOS and overshooting m is not possible. The model can be
interpreted and extended in reasonable ways, however,
To our knowledge, the relevance of chaos theory to the such that chaos becomes a realistic possibility. There are
study of marketing has not been considered in the market- several ways of doing this.
ing literature, yet chaos theory is a relevant consideration. One way to extend the model has been suggested by
This is because nonlinearities are an important feature of Granovetter and Soong (1986). They proposed a repeat
marketing and economic systems (e.g., Forrester 1980; purchase model of product diffusion that includes negative
Arthur 1990; Redmond 1991). For example, Forrester's as well as positive interpersonal influence effects. Their
(1961, 1980) work on system dynamics shows that non- model can be summarized as follows. Two types of inter-
linear phenomena in the form of various kinds of feedback personal effects are proposed, what they term forward and
loops are central to behavior in complex systems such as reverse bandwagon effects. A forward bandwagon effect
marketing systems. Furthermore, the role played by the occurs when a person's purchase is positively correlated
logistic equation in many marketing models is further with those of others. This parallels the positive imitation
testimony to the importance of nonlinear behavior. effect of the Bass model. A reverse bandwagon, negative
In the following section, we briefly review the use of imitation, or "snob" effect occurs when a person's pur-
the logistic equation in models of product diffusion and chase is negatively correlated with the number of existing
market evolution and analyze the conditions under which buyers. These imitation effects depend on two kinds of
complex dynamics, including chaos, can occur. This is thresholds that characterize a person's behavior. The first
followed by a more extensive analysis of a nonlinear or lower threshold is the percentage of people that have to
model of brand competition. purchase the product before an individual will purchase it.
Some people will buy even if no one else has bought (the
Product Diffusion Models innovators), and as the percentage of purchasers increases,
more and more people will buy the product as their thresh-
The logistic equation is an important component of the old is reached. The second or upper threshold is the per-
models of new product diffusion typically used in market- centage of people who have to purchase the product before
an individual will stop buying the product because too
ing (e.g., Mahajan, Muller, and Bass 1990). For example,
many are buying it. For example, a male consumer may
Bass' first purchase diffusion model can be summarized in
not buy narrow ties unless at least 40 percent of his friends
terms of the following equation (Mahajan, Muller, and
have bought narrow ties, but he may switch to some other
Bass 1990, p. 4):
kind of tie in order to be different if 90 percent of his friends
are buying narrow ties.
ANt=Nt-l-Nt=P(m-Nt)+qNt( 1 -Nt/m) [2] Buyers at any time are those whose lower threshold is
exceeded but whose upper threshold is not exceeded. The
wh~e function showing the relation between the percentage of
buyers in one period and the percentage of buyers in the
Nt is the number of adopters in period t
m is the potential number of ultimate adopters next can take on various shapes, depending on the distri-
p is a parameter reflecting the rate of innovation bution of thresholds in the market. A function with an
q is a parameter reflecting the rate of imitation or isolated maximum is possible. This would occur when the
learning. number of buyers grows in the next period for small
percentages of existing buyers (because the forward band-
Two kinds of processes determine the rate of adoption wagon effect dominates) and the number of buyers de-
per period. The first is the rate of innovation, which is clines in the next period for large percentages of existing
reflected in the first part of the equation, p(m - N,). This buyers (because the reverse bandwagon effect eventually
indicates the rate at which people or entities yet to adopt dominates).
 Hibbert and Wilkinson / CHAOS THEORY 221

Equilibrium exists when the percentage of buyers in one lations of organizations (e.g., Carroll 1988) and is summa-
period reproduces the same percentage of buyers in the rized in the following equation: 3
next period, that is, the number of new buyers exactly
balances the number leaving the market. With an isolated dN/dt = rN[(K- N)/K] [31
maximum, such an equilibrium may not occur, as we have
where
seen in the previous section. A sequence of over- and
undershooting of the equilibrium can occur, resulting in N is the number of organizations in the population
cyclical behavior and chaos depending on the nature of the of interest
function. Furthermore, the shape of the resulting function K is the carrying capacity
is likely to be affected by factors such as price and adver- r is the rate of increase and represents the difference
tising. For example, Granovetter and Soong (1986) argued between rate of organizational births and deaths.
that decreases in price are likely to reduce people's lower
thresholds and increase their upper thresholds, and they Unlike the Bass model considered previously, over-
show how price changes can affect the kind of dynamic shooting the carrying capacity is possible, resulting in
regimes that emerge in the market. negative growth (Lambkin and Day 1989, p. 10). This
Another way of introducing chaos is to include interac- introduces the potential for chaos because it means that the
tions between demand and supply in the model. Granovetter equation has the necessary isolated maximum.
and Soong (1986) show how this can be done. Diffusion The Lambkin and Day model is an example of a density-
models, such as that of Bass, generally ignore supply side dependent model of growth. This type of model is com-
effects by implicitly assuming a perfectly elastic supply, monly used in ecological research. A detailed analysis of
the dynamics of this type of model has been provided by
but in a real market suppliers have imperfect information
Guckenheimer, Oster, and Ipaktchi (1977). The emergence
about future demand, so excess supply and demand situ-
of chaos depends on the amount of overshooting that is
ations commonly arise. These situations in turn can lead
possible, and this in turn depends on the value of r (see
suppliers to alter prices as well as other factors that in tum
Appendix A for proof). A stable equilibrium value occurs
affect demand. "[P]rice in the next period is determined by
when r is less than 2, periodic patterns occur for 2 < r <
current price and current excess supply or demand; de-
2.57, and chaos occurs for values of r greater than 2.57.
mand in the next period is determined by consumers'
Whether such values of r can actually occur is debatable.
responses to current demand and by price in the next
In biology, only a few examples of actual populations having
period" (Granovetter and Soong 1986, p. 95). The interac- the necessary growth characteristics have been observed
tion between demand and supply can result in complex (Guckenheimer, Oster, and Ipaktchi 1977, pp. 139-140).
dynamics depending on the response patterns involved. The potential for chaos increases when more complex
Moreover, under some conditions complex dynamics may systems of interacting populations of firms are considered.
arise without the need to assume any upper thresholds or Equation 3 represents the growth of a population of firms
negative interpersonal influence. in isolation and, as noted already, one-dimensional sys-
A third way of introducing chaos is to include interac- tems tend to be more stable as they generally require more
tions among diffusion rates for different products in the extreme parameter values for chaos to emerge. The growth
model. So far the models have considered the diffusion of of one type of firm, however, is likely to interact with the
one product in isolation. "The Bass model assumes that the growth of firms operating in competing and complemen-
adoption of an innovation does not complement, substitute tary markets, resulting in complex feedback effects. The
for, detract from, or enhance the adoption of any other degree of nonlinearity required for such complex systems
innovation (or vice versa)" (Mahajan, Muller, and Bass to become unstable is less extreme.
1990, p. 13). Products do not exist in isolation, however,
and the pattern of diffusion of one product interacts with
those of other products. The growth in demand of competi- AN ATTRACTION MODEL
tive products will tend to inhibit demand, and the growth OF BRAND COMPETITION
of complementary products will tend to accelerate growth. WITH FEEDBACK
Complex patterns of feedback effects can result, and in
such circumstances the potential for chaos increases. In To demonstrate more fully the way chaos theory can be
general, less severe nonlinearity is required for chaos to used to examine the dynamics of nonlinear systems, a
occur in multidimensional systems such as this compared simple dynamic model of brand competition is developed.
to one-dimensional (one-product) systems (Guckenheimer, This is done in order to show the kinds of models that could
Oster, and Ipaktchi 1977). be developed in marketing and how complex dynamics in-
cluding chaos can arise under plausible conditions. The form
Market Evolution Models of the model is quite general and is capable of a wide vari-
ety of applications, including interbrand competition, spa-
Lambkin and Day (1989) have proposed an ecological tial competition, and interfirm competition. It is in part based
model of market evolution using the logistic equation. on models of urban dynamics proposed by Harris and Wilson
Their model is based on models of the evolution of popu- (1978) and Allen and Sanglier (1978, 1979) that are exten-
222 JOURNAL OF THE ACADEMY OF MARKETING SCIENCE S U M M E R 1994

sions of the attraction or gravity model originally proposed zSxt = xj,t - xj,t_1 = f(revenue - costs) =
by Huff (1964) in relation to retail competition. s - xj,t-1)xj,t-1
A brand's market share in a period is assumed to be
proportional to its marketing effort. This is summarized in or
the following equation (the t subscript that indicates the
time period has been omitted for this and subsequent Xj, t = Xj,t_ 1 + ~;(bj,t_ 1 - xj,t_l)Xj,t_ 1 . [6]
equations for the sake of clarity):
The parameter e indicates the degree of response of
a x? marketing effort per period (0 < e < 1). The smaller e is,
the smaller is the degree of response per period for a given
sj = Z a~x~ [41
level of profit. The level of marketing effort in the previous
period also affects the degree of response, such that brands
with smaller levels of marketing effort cannot change as
where much as brands with greater levels. Equation 6 is of the
type commonly used to model growth and decline in many
sj is the market share of brand j in period t
xj is the total marketing effort for brand j in period t types of systems. For small levels of marketing effort, the
aj is a brand-specific attractiveness factor (or a rate of increase or decrease in marketing effort is slow and
brand-specific marketing effectiveness factor) the rate of growth reaches an upper limit as a firm's market
o~ is a parameter reflecting the response of demand share approaches 100 percent. The equation also accords
to changes in marketing effort. with research that indicates that firms' marketing efforts,
such as advertising budgets, are often set in accordance
This type of function is similar to many used in mod- with previous sales levels and profits (e.g., Rossiter and
els of market share, brand, and spatial competition (e.g., Percy 1987).
Black, Ostlund, and Westbrook 1985; Cooper and The equilibrium level of marketing effort for a brand
Nakanishi 1988; Gautschi 1981; Kotler 1984; Naert and occurs when revenue equals cost, or
Weverbergh 1981). For the purposes of this model, mar-
keting effort is assumed to be measured in terms of total bj,t = xj,t. [7]
marketing costs incurred per period. An alternative would
be to develop a more complex attraction function involv- The marketing effort for each brand is determined in-
ing different dimensions of marketing effort and associ- dependently, depending on revenue and costs. Changes in
ated response elasticities, but this would complicate the the marketing effort for a brand affect not only its own
analysis unnecessarily. relative attractiveness and costs but also other brands'
The sales revenue for a brand in a period is equal to the relative attractiveness and their equilibrium level of mar-
total consumer market expenditure in the period multiplied keting effort. Marketing efforts will continue to change
by the brand's market share, or until all brands are in equilibrium. Several questions arise.
Will a general equilibrium emerge, and how stable is it?
bj = Bsj [5] What is the path to equilibrium? Under what conditions, if
any, will the system exhibit complex dynamics such as
wh~e chaotic behavior? Are these conditions likely to exist in the
real world? We next examine the pattern of behavior of
bj is the sales revenue for brand j in period t marketing effort over time under different conditions. We
B is the total consumer market expenditure in period t show how complex dynamics, including chaos, emerge
sj is the market share of brand j. depending on the values of key parameters in the equation.
Critical values are identified that determine the dynamics
The total costs incurred for a brand each period equal of a firm's marketing effort and sales revenue.
their marketing costs. All other costs are assumed constant. The model analyzed here is fairly simple, but as noted
Issues such as economies of scale in marketing effort and previously, it can be extended to incorporate more complex
other costs could be included, but as they do not substan- attraction and cost functions and other feedback effects.
tially affect the nature of the results to be presented, they For example, different elements of marketing effort such
have been excluded. as price and advertising can be modeled separately.
The marketing effort for a brand each period changes
in response to the revenue and costs of the previous period. zSPt= Pj,t - Pj,t-1 = f(revenue - costs) =
If sales revenue exceeds costs such that a profit is made, [8]
marketing effort is increased. If costs exceed revenue, a
~P(bj,t-i - xj,,-1)I'j,t-1

loss is incurred and marketing effort is reduced. The

amount by which marketing effort changes depends on the A A t = Aj, t - A j , t _ 1 = f ( r e v e n u e - c o s t s ) =
degree of profit or loss incurred in the previous period, [9]
EA(bj,t_l - Xj,t_ 1)Aj,t_l
modified by the extent of marketing effort in the previous
period and a parameter indicating the degree of respon- where
siveness of marketing effort per period. This is summa-
rized in the following equation: pj,t is the price of firm j's product in period t
 Hibbert and Wilkinson/ CHAOSTHEORY 223

Ep is the responsiveness of price per period ing effort affect its market share but do not cause other
Aj,t is firm j's advertising in period t brands to change their marketing efforts. The situation
eA is the responsiveness of advertising per period. approximates a perfectly competitive market if all brands
in the market are similarly very small. The equation for the
The degree of responsiveness per period (reflected in model in this case is as follows:
the e parameter) may be different for each aspect of mar-
keting effort, and interactions among them lead to addi- xt+1 = xt + (b - xt)ext [11]
tional complexities. For example, as price is now de-
termined in the model, the market share in volume terms wh~e
for a brand in a period will depend on its relative attrac-
tiveness (including price), whereas sales revenue will de- b = Baxt/S.
pend on sales volume and price.
In this equation, S is the total attractiveness of all brands
The D y n a m i c s of the in the market and is a constant because the focal brand's
Market Competition Model marketing effort and share are assumed to be too small to
noticeably affect it. Further, b is the sales of the brand in a
Results concerning the evolution in time of one and two period, and other symbols are as given above. The dynam-
brands in a single market are presented for three situ- ics of competition are now affected by the elasticity of
ations--a single brand having 100 percent of a market, a demand to changes in marketing effort, ct, as well by the
single brand in a large market such that changes in its own degree of responsiveness of marketing effort per period, E,
level of marketing effort do not noticeably affect the sales and the total size of the market, B.5
of other brands, and two brands competing in the same Equation 11 results in complex dynamics. For the trivial
market. case of t~ = 0, when demand does not respond at all to
Single Brand with changes in marketing effort, the brand's market share is
100 Percent of the Market constant and the equation reduces to the logistic equation
discussed above with the dynamics governed by the prod-
It appears at first sight that the evolution of marketing
uct of market size and the responsiveness of marketing
effort for a brand having all of a market of a fixed size
effort (Be). The addition of ct, the elasticity of demand to
would show no variation at all, but the attraction model
changes in marketing effort, moderates the effect of B and
leads to the following equation:
e. Further, B and e each affect the dynamics of the system,
rather than it being simply the product of these two pa-
xt+~=x t + (B - x t ) ~ t . [10]
rameters. There are values for ct, e, and B that show
This is the logistic equation with r = 1 + BE, as Wilson transitions to chaos, and those for which marketing effort
(1981) has observed.4 The product of e, the responsiveness runs away to infinity or zero. Furthermore, changes in the
of marketing effort per period, and the total size of the value oftx can shift the dynamics from chaos to oscillations
market, B, drive the onset of chaos when 3.57 < 1 + BE < and vice versa.
4.00 or 2.57 < BE < 3.00. Total market size determines the An example of the varied dynamics is seen in the time
path of marketing effort when B = 260, e = 0.01, and ct is
amount of profit or loss resulting at a given scale, and e
indicates the degree of response of marketing effort for a increased from 0 to 0.03. This case was used because e and
given profit or loss. Overshooting and undershooting the t~ have plausible values and BE = 2.6 which, as shown
equilibrium level of marketing effort can occur if market- above, results in chaos in the single-brand case. Marketing
ing effort responds too much per period, resulting in com- effort behaves chaotically until o~ is increased to about
plex dynamics. This means that, for a given degree of 0.019, where a window of three-period oscillations occurs.
responsiveness of marketing effort (e), growth in market As ct is increased further, chaos recurs, followed by a
size over time may result, ceteris paribus, in a shift from a four-period oscillation around ct = 0.024. As ct is increased
smooth adjustment of marketing effort to the changing still further, chaos again results until the values run away
equilibrium to more complex dynamics and chaos. Alter- to zero when ct exceeds 0.03.
natively, for a given market size, changes in the respon- A picture of the dynamic properties of this model is
siveness of components of marketing effort to changes in given in Figure 2. This figure shows the behavior that
revenues and costs, resulting for example from changes in occurs as the size of the market (B) multiplied by the
marketing strategy, may push a brand into or out of regimes response rate (e) varies from 1 to 3 and demand respon-
of complex dynamics. siveness (tx) varies from 0 to 0.1. When ct = 0, behavior
A Small Fish in a Large Pool depends on BE, as shown on the vertical axis. Behavior
Consider the case of a brand with a very small market changes from stable to periodic to chaotic motion and then
share of a large market, such that changes in its level of "explodes," or runs away to infinity or zero. This is the
marketing effort do not noticeably affect the relative attrac- same as the single-brand case. As o~increases, the onset of
tiveness and hence market share of other brands in the periodic behavior and chaos take place at lower values of
market. As a result, changes in the brand's level of market- BE. The values of e and B each matter, not just their
224 JOURNAL OF THE ACADEMY OF MARKETING SCIENCE SUMMER 1994

FIGURE 2
Small Fish Model: Dynamics as a Function of B, ~, and c~

=0.1 = 0.001
1.0

3.0
0.0 0.1 0.0 a 0.1

[--] Stable Period 2 [] Period 4, 8

F I Chaos . . . . . . . . . . Period 3 Explodes

product, as in the case of the single brand. The figure shows Sx = market share of brand x
the situation when the response rate (e) is 0.1 and 0.001. It Sy = market share of brand y.
shows that the onset of periodic behavior and chaos occur
Because there are only two brands competing, the equa-
at lower values of Be the smaller is the value of e. This
tion for market share can be simplified as follows:
means that the less responsive marketing effort is to
changes in profit, for a given value of Be, the more unstable IX IX IX
Sx = x t / ( x t + ~Yt ) [14]
is the system. Of course, to achieve a given value of Be
requires a much greater market size (B) when ~ is much
smaller. This result serves to point out the complexity Sx = (1 -- Sy). [15]
involved in analyzing even quite simple nonlinear models,
in that they do not allow us to introduce the effect of The parameter 13is the ratio of the brand-specific attrac-
additional parameters and variables in a simple additive tiveness ofy to that ofx (i.e., ay/ax). It is in effect a measure
fashion. of the degree of heterogeneity of the brands in the market.
Two Brands Competing in One Market These equations are linked, with x and y affecting each
other's market share. They still maintain the basic form of
Two brands competing in a single market can be repre-
the logistic equation and show a similar type of behavior.
sented by the following two equations:
The equations resemble the predator-prey models that
Xt+l = Xt + (Bsx - Xt)Ext [12]
have been studied extensively in ecology and elsewhere
(e.g., Peschel and Mende 1986). In general, research shows
that two species, in this case brands, can both survive if
Yt+l = Yt + (Bsy - yt)EYt [13] growth is constrained more by internal factors than by
competition. In terms of the brand competition model, the
where
degree of competition is indicated by tx. In the extreme
xt = marketing effort of brand x in period t case of ct = 0, no competition exists. The attractiveness of
yt = marketing effort of brand y in period t each brand is a constant, and the two brands act inde-
 Hibbert and Wilkinson/ CHAOS THEORY 225

pendently with market shares depending on the relative FIGURE 3

brand-specific attractiveness. The situation is equivalent to Examples of the Dynamics
two single-brand markets. The larger c~ is, the more of the Two-Brand Model
changes in one brand's marketing effort affect the market
share of the other. For c~ less than 1, both brands can a) Smooth Paths to Equilibrium
survive. A more formal analysis and illustration of this is
given in Appendix B.
Other parameters of the model now affect the dynamics
of marketing effort in the two-brand case, in addition to
total market size, the responsiveness of marketing effort, /"

and the elasticity of demand to changes in marketing effort.

These are the degree of heterogeneity of the brands as
reflected in 13and the starting levels of marketing effort for
E
the two brands. For the purpose of this analysis, we will
assume that the degree of responsiveness of marketing
effort e and the elasticity of demand to changes in market-
ing effort ~ are the same for both brands. This would be
II [I It I111 II 11 II II il II tl II II II II II I1 I[ II II I111 II II I I H I I m ii II 11 II H II II I[ II II II II 11 II II II II II II II II
yet another source of potential heterogeneity in the market. lime
Study of these equations shows a progression from smooth
transitions to equilibrium values of marketing effort for the
b) Joint Cycles with Delayed Onset
two brands, through oscillations to chaos as the size of the
market (B) increases, as the responsiveness of marketing
effort (a) increases, as the ratio of brand-specific attractive-
ness ([3) deviates from unity, and as the response elasticity
of demand to changes in marketing effort (~) increases.
The starting values for the two brands can affect the i../.....-.............................
dynamics as well.
To illustrate the varied dynamics of this system, three
cases are presented. Their dynamics are shown in Figure 3. E

The cases have been chosen to illustrate particular dynamic

regimes under plausible model conditions.
In Figure 3a, B = 50, ~3= 0.8, ~ = 0.2, and e = 0.01. The
figure shows a smooth path to equilibrium, where the 0 t l ~ " l ! i I I [ l l l t l l H H IErll IItl f l l l l l f I l l l l l l l II I l l L l l l I I I I l l f I l l l IIII Ill II I I l l fll II rl II IH f/II II I I I I I I

equilibrium depends on brand-specific relative attractive- time

ness ([3). Both brands begin with equal marketing effort

and market shares and grow smoothly to their equilibrium c) Joint Chaos
shares and level of marketing effort. The product of B and
is 0.5, which is in the range where stable behavior occurs
in the single-brand and small fish cases.
In Figure 3b, B = 4,000, [3 = 0.8, (x = 0.09, and e = 0.001.
In this situation, the market is much larger, but the degree
of responsiveness (e) is low. The responsiveness of demand
~a
i i
((x) is also lower. The product of B and e is now 4, which .=_

is in the range where chaotic behavior occurs in the single-

E
brand and small fish cases. Now the market is shared
between two brands, and this moderates the amount of
profit or loss made in each period and therefore the degree
of response and overshooting taking place. Joint oscilla-
tions occur after a long period of more stable behavior. it II it II II II I1 II H [I II II [i H 11 [I [] II II IIII I[ I[ II Ill II I III II If II II II H II tl II II ~111 i~ ii It II II II II II II

time
In Figure 3c, B = 50, 13= 0.8, a = 0.2, and E = 0.1. Here
the responsiveness of each brand (e) is greater, but other-
wise the conditions are the same as for the first case. This
results in both brands changing their marketing efforts
chaotically, with periods in which the weaker brand has
greater marketing effort. Note also the window of limited This section has shown how complex dynamics, includ-
fluctuations occurring in the midst of the chaotic swings ing chaos, can occur in some existing nonlinear marketing
in marketing effort and hence market shares. models under plausible conditions. The analysis shows
226 J O U R N A L OF THE A C A D E M Y OF M A R K E T I N G SCIENCE S U M M E R 1994

how sensitive the dynamics are to changes in the values of FIGURE 4

key parameters. In the course of the analysis, we have also Detecting Chaos with Return Maps
illustrated some of the methods that can be used to study
the behavior of nonlinear systems. In the next section, we (a) Chaotic and Random T i m e Series
turn to the ways of detecting the presence of chaos in actual
marketing systems.

:: -.,. . -- :.
DETECTING CHAOS IN ACTUAL = t s a'~l i 9 ~-

MARKETING SYSTEMS ,," = =

,, ; '.
So far we have considered the conditions under which
simple nonlinear equations governing behavior can result
in complex, random-looking behavior. Often, however, the
I I I I I ~ S I I I I I I I I I ~' I I I
problem is that we do not know the rules governing the
behavior of a system. Traditionally, complex, random- time
looking patterns of behavior over time have been ex-
plained in terms of random external shocks or inherent
stochastic processes. In some cases, however, such behav-
ior may be in part caused by chaos resulting from an
underlying nonlinear system. The problem is to distinguish (b) R e t u r n M a p s for Chaotic and Random T i m e Series
between behavior, or that element of behavior, that is
chaotic and that which is the result of complex random
shocks and stochastic processes. [] 9 I1~ O
Traditional linear statistical techniques, such as time
series regression, are not able to distinguish between chaos []
9 []
and randomness. 6 Consequently, resort must be made to + 43 9 [] 9
other techniques. The problem with chaotic data is that the 9 0 9
[]
sensitivity to initial conditions precludes any fitting in the
normal sense using least squares or maximum likelihood
techniques. Thus the parameters of the model set the
pattern (i.e., stability, oscillations, chaos) but the given I DI t I [] I
0
initial conditions determine the particular values in the 0
series. For example, with the same model, parameter val- x(t)
ues, and starting values, different time series can be gen- I O Random 9 Chaos
erated on different computers by virtue of their different
rounding errors. Therefore, the fit of a given time series to
a chaotic model is machine dependent and will give mean-
ingless parameter estimates. During the optimization, a
test series is generated from the current set of parameters Return Maps
and is compared with the actual data. For the same parame-
ters and starting values, each machine will produce a Figure 4a shows two time series, one random and one
different calculated series, so the next guesses of parame- chaotic, generated from Equation 1 with r = 3.99. Visual
ters (which are based on the difference between the calcu- inspection shows no apparent difference between the two,
lated and actual data) will diverge. For real data that are a but if the value at t is plotted against that at t + 1, which is
combination of chaotic, random, and nonchaotic elements, referred to as a return map, a clear difference emerges. For
it will be possible to fit the nonchaotic components. The the chaotic series in Figure 4b, the outline of a parabola
remaining chaos and random elements will be treated as emerges, whereas for the random series no pattern
random errors. How this compromises a given optimiza- emerges. Return maps can also be plotted for the values of
tion will depend on the relative magnitudes of the different t against t + 2 or t + 3 and so on, and in each case the chaotic
components of the time series. series reveals a hidden pattern, whereas random behavior
Four basic types of methods to detect chaos exist and does not. The shape of the resulting pattern depends on the
are described below. These are, respectively, return maps, nature of the underlying equations. In Figure 4b, the pa-
the correlation dimension, the Lyapunov exponent, and rabola results because the series is generated from the
prediction error. For a more detailed review of the first logistic equation. Even if the system under investigation is
three methods see Brock (1986) or Barnett and Choi multidimensional and only one aspect of it is measured
(1989). over time, a return map will indicate if there is evidence of
 Hibbert and Wilkinson / CHAOS THEORY 227

chaos. If the system is complex, a large number of data The Lyapunov Exponent
points may be required before any pattern becomes clear
to visual inspection. The method described next attempts The Lyapunov exponent is a measure of the dynamic
to detect the presence of such a pattern. properties of a system and is developed as follows. A
property of a chaotic system is that points that initially are
Correlation Dimension close together separate exponentially. After N steps, the
separation of two points that initially differ by e is
The correlation dimension is a measure of nonrandom- exp(N)~), where ~,is known as the Lyapunov exponent. For
ness in return maps. It was developed by Grassberger and chaotic motion ~. is positive, for a stable equilibrium point
Procaccia (1983) and focuses on the way, points repre- )~ is negative (i.e., points that are far apart come together
senting sequences of values are scattered in m-space, such at equilibrium), and for periodic oscillation )~ is zero. The
as the way sequences of three values are scattered in a magnitude of a positive Lyapunov exponent will also
three-dimensional space with axes t, t + 1, and t + 2. If a determine the speed of divergence of two paths. For exam-
data series is random, the points will be randomly scattered
ple, if )~ = 0.693, the separation between two paths will
throughout m-space, but if it is chaotic, they will cluster
double on each iteration.
more. The correlation dimension is a statistic that measures
The Lyapunov exponent can be derived from the equa-
the way the m-space is filled by data points and has been
tion governing a system. It can also be estimated directly
widely used to determine the presence of chaos. The pro-
from time series data by the method of Wolf et al. (1985)
cedure for calculating its value is in essence straightfor-
as follows. First, plot the data points in m-space, where m
ward, although it is complex to implement. It can be sum-
is a sequence of m-time periods Xt, St+T, Xt+2T. . . . . Xt+(m_l)T.
marized as follows. First, plot the data points in m-space,
Next, take the first point in the data series and find the
where m is a sequence of m-time periods (i.e., Xt, Xt+T, distance to its nearest neighbor L(to) (the point with the
Xt+2T. . . . , Xt+(m 1)T), For example, if m = 3 plot all se- most similar series of values over m-time periods). Third,
quences of three values in a three-dimensional space t, t +
observe the two points as they evolve and compute the
1, and t + 2. Next, for each point in the m-space, count how
distance apart, referred to as L'(tl), after a short time. The
many points are within a small absolute distance of it in
time is chosen to allow only a small change in the value of
any direction (aim). Increase the distance away by a small
the variable. In many cases in which discrete data are
amount and count how many more points are now within available, this will be one time period. Fourth, find a new
that distance, and repeat this for greater and greater dis- point L(tl), repeat the third step, and continue until the end
tances. Finally, calculate the correlation dimension (C~m): of the data set. Finally, calculate the Lyapunov exponent
(L) as follows:
VlnCm(C)]
O~m=E-~=lim L lnc J [16] M
L'(tk)
s 1 E l~ - - [181
wh~e tm - to L(tk- 1)
k=l

lim {(i,j) I am- a~ I < e} where M is the total number of sets of points examined.
fro(e) = [17]
Prediction Error
Nm=N- ( m - 1) This method measures the accuracy of predictions
based on the past behavior of a data series. It was proposed
Cm(F~) is the number of m-points that are closer together by May (1990) and is based on the premise that in chaos
than e, and C~mis the dimension of this filling of m-space. one can make short-term, but not long-term, predictions.
The notation in the { } braces means to count the number In pure random systems, one cannot even make short-term
of occurrences for which the condition is satisfied. predictions. The procedure is as follows. First, take every
This method also can be used to estimate the dimen- series of m-points from a data series with m = 2. Second,
sionality of the chaotic system, or the approximate number take the first set of m-points and find its nearest neighbor.
of degrees of freedom in the underlying system of equa- Third, determine the error in the m + lth point in the series,
tions necessary to produce the chaotic series. If aj goes to where error is defined as the euclidian distance between
some small value as (xm becomes large, the limiting (Xmis the actual and nearest neighbor's values. Fourth, repeat the
a good indicator of the dimensionality of the chaotic sys- process for each series of m-points in the data series. Fifth,
tem. It should be noted, though, that the size of the data set compute the average error. Sixth, repeat the first five steps
constrains the maximum size of the correlation dimension with increasing values of m. Finally, graph the average
that can be calculated to 21ogl0(N). Hence, for a data set of error against m.
1,000 points, the maximum dimension possible is six. If The presence of chaos is indicated by good predict-
the underlying system is a higher dimensional system, it ability (low average error) when m is small but worse
cannot be detected. The method thus places severe de- predictability as m increases. Figure 5 shows the average
mands on the size of the data series required. prediction error for the random and chaotic data series
228 JOURNAL OF THE ACADEMY OF MARKETING SCIENCE SUMMER 1994

FIGURE 5 there is no reason to think consumer goods markets do not

Prediction Error Test: Chaotic also. The problem is to identify situations in which such
versus Random Time Series an explanation is likely to be relevant. As we have shown,
techniques have been developed to do this, although they
1 require long data series of high quality in order to reliably
identify chaotic behavior and nonlinearities.
0.1
Attempts have been made to identify chaos in various
economic and financial time series such as business cycles,
monetary aggregates, stock prices, and employment rates.
0.01
These have met with some success, although researchers
have been hampered by the lack of sufficient high-quality
0.001 data (Brock and Sayers 1988; Grandmont and Malgrange
1986; Barnett and Choi 1989; Hinich and Patterson 1989).
0.00Ol In marketing, the advent of scanner data has resulted in the
generation of rich and reliable time-series data eminently
0.00001 1 I suitable for this type of analysis. The task, however, will
0 10 2O 30
not be easy.
Delay
The first type of problem is that competitors are likely
I . Chaos -B Random~ to change their rules of behavior in response to changing
market conditions, including the introduction of new
brands and the removal of existing ones. Second, there are
problems associated with separating out different compo-
shown in Figure 4a. The chaotic series shows low error for nents of a time series. Traditional techniques for isolating
small m, whereas for the random series the average error the effects of external factors and stochastic processes,
is approximately constant. such as regression and smoothing techniques, may inad-
In chaos, the average error increases with m because as vertently remove elements of chaotic motion from a time
m increases, data from increasingly more distant time series so that it can no longer be easily detected. This
periods are being used to select the nearest neighbor. This problem calls for additional research in its own right, as
introduces more error into the prediction. Data series that little is known about the robustness of different chaos
are most similar for two periods are more similar in the detection techniques or how the presence of chaos along
next period than data series that are most similar for longer with other components affects regression estimates (Gordon
periods. 1991). Despite these problems, time windows are likely to
be detected in which chaos and other complex dynamics
occur for a limited period.
IMPLICATIONS OF CHAOS THEORY An important question is whether the parameter values
required to produce chaos will occur in actual marketing
Implications for Marketing Theory systems. When simple dynamic models are analyzed,
chaos theory may appear to be largely irrelevant because
Chaos theory offers important implications for the the parameter values required for chaos are too extreme
study of marketing systems. We have shown how complex and unrealistic. This is not the case here, where plausible
dynamics can arise from relatively simple models of the values of the key parameters of the brand competition
type widely used in modeling aspects of marketing. Criti- model lead to complex dynamics. The responsiveness of
cal parameter values have been shown to exist that are elements of marketing effort to changes in brand shares
associated with shifts from smooth paths to equilibrium to and profit levels is particularly important. This affects the
regions exhibiting oscillations of varying complexity to likelihood of brands overshooting equilibrium levels of
chaotic regimes. These dynamic patterns are completely marketing effort, which is a key factor determining chaos.
overlooked by traditional analytical techniques that focus For some aspects of marketing effort, the speed of response
on equilibrium conditions and local stability. Several is- will be quite slow. Advertising budgets and prices, for
example, are set weeks and months ahead and cannot be
sues arise.
changed every day or week. This would tend to have a
Does Chaos Existin Marketing Systems? dampening effect on the system or require long time peri-
Chaos theory offers an alternative way of explaining the ods before any chaotic pattern reveals itself. However,
kind of complex, random-looking patterns of behavior other aspects of marketing effort may be more responsive,
often found in marketing time series such as sales, inven- such as salesperson call patterns or in-store discounting.
tory, brand shares, and profits (e.g., Bishop, Graham, and Fortunately, it is not necessary to know the mechanism
Jones 1984; Preismeyer and Batik 1989). Chaos theory driving complex dynamics. The chaos detection tech-
shows that such complex, random-looking behavior can niques described here search for the hallmarks of chaos in
arise endogenously in a deterministic way when non- the behavior resulting from these mechanisms, in our case
linearities exist in system behavior. Weather patterns, deer embedded in the pattern of behavior of brand sales and
populations, and stock prices seem to display chaos, so market shares over time. Once indicators of chaos are
 Hibbert and Wilkinson / CHAOS THEORY 229

discovered, a search can be made for the underlying patterns of segmentation and differentiation that tend to
mechanisms at work. In this way, researchers' efforts are reduce the possibility of chaos. Any divergence from this
directed to areas where detailed investigation is more pattern may trigger the onset of complex dynamics includ-
likely to be fruitful, rather than wasting time investigating ing chaos, which will drive the system back to the original
time series that only contain random behavior. This is the pattern or to another pattern that reduces the chance of
approach taken in other areas, such as the search for chaotic chaos. Periods of chaos and oscillations may represent,
behavior in stock market behavior or in the variations in therefore, important transition stages to new, stable market
species numbers over time. structures and patterns of competition, with chaos playing
As a final point, it should be remembered that real a role in speeding the pace of change and "testing" the
marketing systems are not simple one-dimensional sys- viability of alternative structures.
tems. They are multidimensional systems that involve The work of Medio (1991) suggests another way in
complex sets of interactions. This, as has been noted, which the pattern of differentiation existing among system
increases the potential for chaos because the degree of members affects the likelihood of chaos. He demonstrated
nonlinearity required is less severe in higher-dimensional that chaos can arise as a result of the combination of
systems (May 1976; Guckenheimer, Oster, and Ipaktchi nonlinearities and the pattern of response lags existing in
1977; Granovetter and Soong 1986). The analysis pre- a system among individual members. Medio showed that
sented above illustrates this. Chaos is a more common low variance in the distribution of individual lags, which
occurrence when more than one brand is considered than implies a more rigid, homogeneous type of response to
when the market consists of a single brand. stimuli, tends to produce chaotic behavior. Such rigid
Chaos and Structural behavior makes the dampening of fluctuating system be-
Change Mechanisms havior more difficult. If responses are too uniform, the
Chaos theory can also contribute to understanding of resulting complex dynamics and chaos create incentives to
structural change and evolution in marketing systems. In change behavior strategies and/or leave the market. This
the brand competition model explored in this article, the will continue until a more heterogeneous and stable pattern
rules governing the determination of marketing effort and of responses emerges.
demand response are fixed. In reality, as noted previously, The type of explanation of structural change described
these rules may change in response to the behavior taking in the foregoing has been proposed by Day and Walter
place in the market. Firms with brands operating in chaotic (1989) as a way of understanding economic growth as a
regimes would find it difficult to plan effectively, and this multiphase dynamic process in which shifts occur between
will create incentives to avoid or minimize the occurrence different technologies and associated infrastructures.
of chaos. The problem is complex because, as we have Chaos theory helps explain the way in which these shifts
seen, firms may not be able to recognize or control the occur. The approach is also the basis for the dynamic
onset of chaos through the control of their individual theory of the evolution of marketing systems' structures
brand's marketing effort. Instead, more general feedback proposed by Wilkinson (1988, 1990).
and evolutionary mechanisms may be envisaged that mod- Another type of resulting structural response is sug-
ify the rules by which firms alter the marketing effort for gested by the work of Day (1983a). He proposed that chaos
brands and/or the rules governing the way demand re- may result in the evolution of certain types of economic
sponds to changes in marketing effort. Over time, firms institutions and practices that serve as "disequilibrium"
and brands may exit and enter the market until populations mechanisms. These mechanisms buffer and ameliorate the
of brands and firms emerge that behave in such a way that effects of unpredictable fluctuations in behavior and allow
chaotic regimes are minimized or avoided. a system to function in disequilibrium. Day argued that
Some possible examples are suggested by the brand such mechanisms include inventory-holding intermediar-
competition model. We have shown that in a one-brand ies such as retail stores, banks and financial intermediaries
market, the onset of chaos depends on the size of the that regulate the flow of purchasing power, ordering
market (B) and the responsiveness of the firm (e). A mechanisms with accompanying backlogs and variable
feedback mechanism may occur such that over time firms delivery delays, and insurance systems.
learn to tone down the responsiveness of their marketing
efforts for brands as the market expands, by reducing e as Management Implications
B increases. This would reduce the likelihood of chaos
occurring. Our analysis shows why it may be very difficult to make
In multibrand markets, the onset of chaos also depends predictions based on a profile of past behavior. In chaotic
on the degree of competitive interaction (~) among the regimes, even when the rules of behavior are known pre-
brands. This will be affected by the amount of segmenta- cisely, the results are extremely sensitive to starting values.
tion and product differentiation practiced by firms in the This means that while the next point in a time series can
market. As segmentation and differentiation increase, they be calculated, errors in the calculation of subsequent points
will tend to isolate brands somewhat from direct competi- compound, making long-term prediction practically im-
tion with other brands and hence reduce the impact of possible.
changes in marketing effort on relative attractiveness (i.e., Moreover, as Day (1983a) argued, the past behavior of
reduce the value of ~x). Over time, this could result in stable a system capable of chaotic dynamics "may be a poor guide
230 JOURNALOF THE ACADEMYOF MARKETINGSCIENCE SUMMER 1994

for inferring even qualitative let alone quantitative pat- effort and marketing effort responds to brand sales. The
terns of change in the future" (p. 135). Small changes in challenge now is to examine market time-series data to
key parameters can result in shifts from one dynamic detect the presence of chaos and other dynamics arising as
regime to another, and long periods of apparently stable a result of nonlinear behavior systems and to distinguish
behavior can take place before the onset of fluctuations such behavior from purely random behavior and the im-
(e.g., Figure 3b). Furthermore, in chaotic regimes, periods pact of exogenous factors. We have described some of the
of moderate fluctuations can occur followed by the occur- techniques for doing this. Once having detected the pres-
rence of larger fluctuations (e.g., Figure 3c). Such behavior ence of such dynamics, more detailed investigation of the
could lead to false conclusions that a system is stable or underlying mechanism can be undertaken to identify the
has settled down, especially if the time ilavolved is a matter processes causing it. This task parallels work in other
of months and years rather than hours and days. disciplines, where considerable effort has been devoted to
Our analysis shows that the critical parameters driving examining time-series data to detect the presence of non-
the dynamics of the system are not necessarily in the linear motion. As yet, despite the use of sophisticated
control of individual competitors. For example, in the techniques, only weak evidence of chaos exists. This has
brand competition model developed here, more general not stopped researchers from continuing to work in the
system properties such as market size, the degree of com- area.
petitive interaction, and the responsiveness of competitors A second direction for future work is to reexamine the
are critical. In such circumstances, solutions to the prob- many nonlinear dynamic models used in marketing in
lems of chaos are not to be found at the individual brand terms of chaos theory to see to what extent they have the
level. One brand altering its behavior will not alter the potential to exhibit chaotic dynamics under realistic con-
character of the general market system. ditions. This article has demonstrated how this can be
Sometimes the parameters are under more direct con- done. Possible candidates are various dynamic models of
trol. For example, in the diffusion models the rate of consumer/buyer behavior such as those in which behavior
adoption and imitation may be partly influenced by firms' in one time period affects behavior in the next period,
marketing activities. For the brand competition model, the which Benhabib and Day (1981) have shown can result in
key parameters are partly under the control of economic chaos. Models of diffusion and adoption, such as those
regulatory bodies. These can affect general market condi- described above, also warrant attention. Other possible
tions and all brands' behavior, which in turn affect system candidates include production-inventory system models in
dynamics. Without detailed knowledge of the existing which orders and stocks respond indirectly and with lags
market situation and the critical parameter values in- to customers' orders (e.g., Forrester 1961; Bowersox et al.
volved, such intervention could make things worse instead 1972), competitive response models in which firms re-
of better. Without precise controls and knowledge of the spond to other firms' strategies (e.g., Lambkin and Day
system, intervention could move the system from one 1989), and models in which demand and supply respond
region of chaos to another or from simple to complex to each other over time (e.g., Granovetter and Soong 1986;
oscillating regimes and even to chaos. Redmond 1991).
Given the problems in predicting and controlling be- A third direction for research is to extend the brand
havior in chaotic regimes, other types of strategies are competition-type models described here to encompass
called for to deal with the situation. One possibility is the more complex and realistic market situations. Most of the
development of various types of disequilibrium mecha- analysis reported in this article has been done using spread-
nisms to absorb or dampen the fluctuations, such as those sheets to simulate the behavior of the models over time
proposed by Day (1983a). This, of course, runs counter to under varying conditions. The reader can therefore easily
trends toward just-in-time systems that seek to reduce verify and extend the analysis presented. 7
buffer stocks. Such systems involve closer cooperation A first type of extension is the examination of competi-
among the firms involved. This may change their behavior tive dynamics under more complex situations, such as
in such a way as to reduce any over- and undershooting of more than two brands and when marketing effort is decom-
equilibrium supply levels and thereby prevent complex posed into elements with different response rates. A second
oscillations and chaos. The tighter coupling that is in- type of extension is to recognize that the response patterns
volved, however, could also lead to instabilities elsewhere (i.e., the rules of behavior) will vary across brands and over
in the system. time. For example, some brands may change their market-
ing effort more rapidly in response to changing sales than
will others. Furthermore, firms may alter the way brands
CONCLUSIONS AND FUTURE respond as a result of changing conditions in the market;
RESEARCH DIRECTIONS this includes brands entering and leaving the market. As
presently constructed, the model assumes fixed identical
In this article, we have shown how complex dynamics behavior rules with brands differing only in terms of their
can arise from a simple dynamic model of market compe- starting position and inherent attractiveness.
tition and that this can have important implications for Models that allow for the rules governing system be-
marketing theory and practice. We have shown that, under havior to evolve in response to system behavior are known
plausible conditions, complex dynamics can arise in a as evolutionary models. These have only recently begun
market in which brand sales depend on share of marketing to be developed as a result of the advent of modem corn-
 Hibbert and Wilkinson / CHAOS THEORY 231

puter technology and simulation techniques such as ge- APPENDIX B

netic algorithms and classifier systems (e.g., H o l l a n d and Equilibrium and Stability of the
Miller 1991; R u t h v e n 1993). Intriguing questions arise, Two-Brand Attraction Model Equations
such as the possible evolutionary paths market systems can
take and what types of stable states exist in terms of The number of variables in the equations can be reduced
response patterns a m o n g brands. U n d e r what conditions somewhat if we write Equation 12 as xt§1 = xt([1 + B~x] - ext)
are brands that change their response patterns rapidly as a and use the substitution X = xU(1 + Besx). The results are as
result of c h a n g i n g market conditions more likely to survive follows:
than m o r e conservative ones? Are there stable mixes of
Xt+ 1 = (1 + Besx)Xt(1 - Xt).
slow and rapidly responding brands?
E v o l u t i o n a r y - t y p e models show that chaotic d y n a m i c s For Equation 13, use the substitution Y= yr + Br which
can play a role in g u i d i n g a system along particular evolu- results in
tionary paths and that m i n o r variations in behavior at
lit+1 = (1 + Br - Yt)'
critical times can have major impacts on future develop-
ment. N e l s o n and W i n t e r (1982), Axelrod (1984), and Day Because s is homogeneous in x and y, any scaling of x and y will
and Walter (1989) provided examples of this type of work. not affect it.
In marketing, a discussion of these types of models and the The stationary state for a given set of initial parameters occurs
role they can play can be f o u n d in W i l k i n s o n (1990), and when xt+1 = x t. For Equations 12 and 13, the condition that xt+1 =
some work has b e e n done to e x a m i n e the evolution of the x t and Yt+l = Yt are fulfilled by (Bs x - xt)r x t = 0 and (Bsy - yt)r
brand competition-type models described above using ge- Yt = 0. Thus both have the trivial solution x t = 0 and Yt = 0, and in
netic algorithms (Hibbert and W i l k i n s o n 1991). There is addition the following relationships hold, where the subscript e
g r o w i n g awareness and interest in these types o f models denotes equilibrium:
in m a n y disciplines. Their role and value in marketing has
y(x e) = xe(1/[3[B/x e - 1]) l&
only j u s t b e g u n to be considered.
We have attempted to demonstrate that chaos is not just X(Ye) = ye(~[B/y.e - 1 ] ) 1/~

a quirky kind of d y n a m i c that needs attention but is part of These two equations show the different values of y for an
a larger focus on d y n a m i c s and evolutionary processes in equilibrium value of x and the different values of x for an
marketing systems that has b e c o m e possible since the equilibrium value ofy. The stationary points (xv y,) occur when
advent of m o d e r n computer techniques. The majority of both x and y are in equilibrium. They are determined from the
marketing scholars tend too easily to rely on traditional crossing points on plots of these equations, such thaty(x~) = x(yr
techniques and models and resist seeing the opportunities The stability of a stationary point is determined from the local
arising from these n e w developments. We hope this article behavior of the functions. If the point is an attractor (that is, after
contributes to legitimizing and accelerating the investiga- small perturbations from the point, the behavior returns to equi-
tion of these other approaches. librium) and both x, and y, are greater than zero, then both brands
survive in equilibrium. For cz >_ 1, both brands cannot simulta-
neously survive in equilibrium. For values of c~ < 1, both brands
can survive.

ACKNOWLEDGMENTS
APPENDIX A
Dynamic Properties of the Lambkin
The research reported is in part supported by an Australian
and Day Market Evolution Model
Research Council research grant for 1990-1991. The authors
Equation 3 can be rewritten as follows. Letx = N/K such that would like to acknowledge the helpful comments of the editor
N is expressed in "carrying capacity" units, which means that xt and the anonymous referees.
varies between 0 and 1. Equation 3 becomes
xt+ 1 - x t = rxt(1 -xt) = NOTES
S t + 1 = Xt + rxt (1 - x t ) = xt + r x t - rx~-- (1 +r) x t - rx~ .
1. Any equation of the form xt+1 = x t + (A - B x t ) x t may be
Let Yt = r/(1 + r)xt, then reduced to it by the substitution Xt = xtB/(1 + A) and the
Yt+ 1 (1 + r)/r = lit (1 + r)2/r - r Yt2 (1 + r)2/r 2 equation becomes Xt+1 = (1 + A)Xt(1 - Xt), where 0 < Xt < 1.
This is the logistic equation for r = (1 + A).
Yt+l = (1 + r ) Y t - ( 1 +r)Y~ 2. For a more detailed explanation and illustration of chaos
= (1 + r)Yt(1 - Yt) theory, the reader should consult one of the excellent re-
views that has appeared. For a less technical introduc-
where Yt varies between 0 and r/(1 + r). tion, a good source including computer programs is Gordon
This equation is in the same form as Equation 1, and the and Greenspan (1988). For a more general discussion see
dynamics depend on the value of (1 + r). Prigogine and Stengers (1984) or Gleick (1987). For a more
232 JOURNAL OF THE ACADEMY OF MARKETING SCIENCE SUMMER 1994

technical i n t r o d u c t i o n see, for e x a m p l e , May (1976), Clarke, Martin and Alan G. Wilson. 1983. "The Dynamics of Urban
Guckenheimer, Oster, and Ipaktchi (1977), or Stutzer (1980). Spatial Structure: Progress and Problems." Journal of Regional Sci-
3. These models owe their origin to the classic Lotke-Volterra ence 23 (February): 1-18.
equations for animal populations developed in the 1920s. Cooper, Lee G. and Maseo Nakanishi. t988. Market-share Analysis:
4. The equation can be rewritten in canonical (universal) form Evaluating Competitive Marketing Effectiveness. Boston: Kluwer Aca-
using the substitution indicated in Note 1. It becomes Xt+ 1 = demic Publishers 9
(1 + bE)Xt(1 - Xt), with X t = xtke/(1 + be). Day, Richard H. 1982. "Irregular Growth Cycles 9American Economic
5. The m o d e l is equivalent to a single brand with all the Review 72 (June): 406-414.
market b u t where its marketing effort affects the size of the . 1983a. "Adaptive Economics." In Cities and Regions as Non-
total market. In both cases, d e m a n d is not constant. linear Decision Systems. Ed. Richard W. Crosby. Boulder, CO:
6. Gordon (1991) has s h o w n that traditional regression tech- Westview, 29-62 9
niques can be useful in determining the lags in system 91983b. "The Emergence of Chaos from Classical Economic
dynamics that may underlie chaotic behavior. This still Growth." Quarterly Journal of Economics 97 (2): 201-213.
leaves unspecified the mechanism producing the behavior. Day, Richard H. and Jean-Luc Walter. 1989. "Economic Growth in the
7. Copies of the Lotus 123 spreadsheets used are available on Very Long Run: On the Multiple-phase Interaction of Population,
request. Technology, and Social Infrastructure." In Economic Complexity:
Chaos, Sunspots, Bubbles and Nonlinearity. Eds. William A. Barnett,
John Geweke, and Karl Shell. Cambridge, MA: Cambridge University
REFERENCES Press, 253-290.
Feigenbaum, Mitchell J. 1979. "Universal Metric Properties of Non-
linear Transformation." Journal ofStatisticalPhysics 21 (June): 669-706.
Allen, Peter M. and Michele Sanglier. 1978. "Dynamic Model of Urban
Growth." Journal of Social and Biological Structures 1 (July): 265- 91983. "Universal Behavior of Nonlinear Systems." Physica D,
280. Nonlinear Phenomena 7 (May): 16-39.
Allen, Peter M. and Michele Sanglier. 1979. "A Dynamic Model of Forrester, Jay W. 1961. Industrial Dynamics. Cambridge, MA: MIT
Growth in a Central Place System." Geographical Analysis 1I (July): Press.
256-272. 91980. "System Dynamics---Future Opportunities." In System
Arthur, W. Brian. 1990. "Positive Feedbacks in the Economy." Scientific Dynamics: TIMS Studies in the Management Sciences. Eds. Augusto
American 263 (February): 80-85. A. Legasto Jr., Jay W. Forrester, and James M. Lyneis. Amsterdam:
Axelrod, Robert. 1984. The Evolution of Cooperation. New York: Basic North-Holland, 7-21.
Books. Gautschi, David A., Jr. 1981. "Specification of Patronage Models for
Barnett, William A. and Seungmook S. Choi. 1989. "A Comparison Retail Center Choice." Journal of Marketing Research 18 (May):
between the Conventional Econometric Approach to Structural Infer- 162-174.
ence and the Nonparametric Chaotic Attractor Approach." In Eco- Gleick, James. 1987. Chaos: Making a New Science. New York: Penguin 9
nomic Complexity: Chaos Sunspots, Bubbles and Nonlinearity. Eds. Gordon, Theodore J. 199 I. "Notes on Forecasting a Chaotic Series Using
William A. Barnett, John Geweke, and Karl Shell. Cambridge, MA: Regression." Technological Forecasting and Social Change 39 (July):
Cambridge University Press, 141-212. 337-348.
Beaumont, John R., Marcus Clarke, and Alan G. Wilson. 1981. "The Gordon, Theodore J. and David Greenspan. 1988. "Chaos and Fractals:
Dynamics of Urban Spatial Structure: Some Exploratory Results Using New Tools for Technological and Social Forecasting." Technological
Difference Equations and Bifurcation Theory." Environment and Plan- Forecasting and Social Change 34 (August): 1-25.
ningA 13 (December): 1473-1483. Grandmont, Jean-Michel and Pierre Malgrange. 1986. "Nonlinear Eco-
Benhabib, Jess and Richard H. Day. 1981 9"Rational Choice and Erratic nomic Dynamics: Introduction." Journal of Economic Theory 40 (Oc-
Behavior." Review of Economic Studies 48 (July): 459-471. tober): 3-12.
Bishop, William S., John L. Graham, and Michael H. Jones. 1984. Granovetter, Mark and Roland Soong. 1986. "Threshold Models of
"Volatility of Derived Demand in Industrial Markets and Its Manage- Interpersonal Effects in Consumer Demand." Journal of Economic
ment Implications." Journal of Marketing 48 (Fall): 95-103. Behavior and Organization 7 (March): 83-99.
Black, William C., Lyman E. Ostlund, and Robert A. Westbrook. 1985. Grassberger, Peter and Itamar Procaccia. 1983. "Measuring the Strange-
"Spatial Models in an Intrabrand Context." Journal of Marketing 49 ness of Strange Attractors." Physics D, Nonlinear Phenomena 9 (Oc-
(Summer): 106-113. tober): 189-208.
Bowersox, Donald J., Omar K. Helferich, Edward J. Marien, Peter Guckenheimer, John, George Oster, and A. Ipaktchi. 1977. "The Dynam-
Gilmour, Michael L. Lawrence, Fred W. Morgan, and Richard T. ics of Density Dependent Population Models." Journal of Mathemati-
Rogers. 1972. Dyncanic Simulation of Physical Distribution Systems. cal Biology 4 (2): 101-147.
East Lansing: Michigan State University, Division of Research, Gradu- Harris, Brian and Alan G. Wilson. 1978. "Equilibrium Values and Dy-
ate School of Business Administration. namics of Attractiveness Terms in Production-constrained Spatial-
Brock, William A. 1986. "Distinguishing Random and Deterministic interaction Models." Environment and Planning 10 (April): 371-
Systems." Journal of Economic Theory 40 (October): 168-195. 388.
Brock, William A. and Chera L. Sayers. 1988. "Is the Business Cycle Hibbert, D. Brynn and Ian F. Wilkinson 91991. "Modelling the Evolution
Characterized by Deterministic Chaos?" Journal of Monetary Eco- of Marketing Systems Using Genetic Algorithms." Paper presented at
nomics 22 (July): 71-90. the 7th Industrial Marketing and Purchasing Conference, University of
Campbell, David K. and Harvey Rose. 1983. Order in Chaos. Amster- Uppsala, Sweden.
dam: North-Holland. Hinich, Melvin J 9and Douglas M. Patterson. 1989. "Evidence of Non-
Carroll, Glen R. 1988. Ecological Models of Organizations. Cambridge, linearity in the Trade-by-Trade Stock Market Return Generating Pro-
MA: Ballinger. cess." In Economic Complexity: Chaos, Sunspots, Bubbles and
Chiarella, Carl. 1988. "The Cobweb Model: Its Instability and the Onset Nonlinearity. Eds. William A. Barnett, John Geweke, and Karl Shell.
of Chaos." Economic Modelling 5 (October): 377-384. Cambridge, MA: Cambridge University Press, 383-409.
 Hibbert and Wilkinson / CHAOS THEORY 233

Holland, John. H. and John H. Miller 91991. "Artificial Adaptive Agents Ruthven, Russell. 1993. "Adapting to Complexity." Scientific American
in Economic Theory." American Economic Review 81 (May): 365- 268 (January): 130-140.
370. Stutzer, Michael. 1980. "Chaotic Dynamics and Bifurcations in a Macro
Huff, David L. 1964. "Defining and Estimating a Trading Area." Journal Model." Journal of Economic Dynamics and Control 2 (November):
of Marketing 28 (July): 34-38. 353-376.
Kotler, Philip. 1984. Marketing Management Analysis Planning and Weiss, Gary. 1992. "Chaos Hits Wall St.--The Theory That Is." Business
Control. Englewood Cliffs, NJ: Prentice-Hall. Week (November 2): 58-60.
Lambkin, Mary and George S. Day. 1989 9"Evolutionary Processes in Wilkinson, lan E 1988. "The Dynamics of Channel Structure and the
Competitive Markets: Beyond the Product Life Cycle," Journal of Structure of Dynamics." Paper presented at the Winter Educators'
Marketing 53 (July): 4-20. Conference of the American Marketing Association, San Diego.
Lundquist, Stig, Norman H. Marsh, and Mario P. Toni. 1988. Order and 91990. "Toward a Theory of Structural Change and Evolution in
Chaos in Nonlinear Physical Systems 9New York: Plenum 9 Marketing Channels." Journal of Macromarketing 10 (Fall): 18-46.
Mahajan, Vijay, Eitan Muller, and Frank M. Bass. 1990. "New Product Wilson, Alan G. 1981 Geography and the Environment: Systems Analytic
Diffusion Models in Marketing: A Review and Directions for Re- Methods. London: Wiley.
search." Journal of Marketing 54 (January): 1-26. Wolf, Alan, Jack B. Swift, Harry L. Swinney, and John A. Vastano. 1985.
May, Robert M. 1976. "Simple Mathematical Models with Very Compli- "Determining Lyapunov Exponents from a Time Series." Physica D,
cated Dynamics." Nature 261 (June 10): 459-467. Nonlinear Phenomena 16 (July): 285-317.
91990. "Chaos and Some of Its Implications in the Biological
Sciences." Paper presented at the International Conference on Chaos
in Australia. Kensington, Australia: University of New South Wales.
Medio, Alfredo. 1991. "Discrete and Continuous-Time Models of Cha-
otic Dynamics in Economics." Structural Change and Economic Dy- ABOUT THE AUTHORS
namics 2 (June): 99-118.
Medio, Alfredo and Giampaolo Gallo. 1993. Chaotic Dynamics: Theory B r y n n H i b b e r t is head of the School of Chemistry at the Uni-
and Applications to Economics. Cambridge, UK: Cambridge Univer- versity of New South Wales. He arrived from London University
sity Press. to take the chair of Analytical Chemistry in 1987 with a back-
Naert, Philippe and Marcel Weverbergh. 1981. "On the Prediction Power
ground in physical chemistry. His interest in chaos stems from
of Market Share Attraction Models." Journal of Marketing Research
work on fractals--those wiggly shapes that appear to be so
18 (May): 146-153.
Nelson, Richard R. and Sidney G. Winter. 1982. An Evolutionary Theory ubiquitous in nature. A chance meeting with Professor Wilkinson
of Economic Change. Cambridge, MA: Belknap Press of Harvard allowed him to use his computing skills in an unusual (for him)
University Press. area of research.
Nichols, Nancy A. 1993 "Efficient?
9 Chaotic? What Is the New Finance?"
Harvard Business Review 71 (March-April): 50-60. Ian F. Wilkinson assumed the Foundation Chair in Marketing at
Peschel, Manfred and Werner Mende. 1986. The Predator-Prey Model. the University of Western Sydney, Nepean in 1991. He has been
New York: Springer-Verlag. a visiting professor at various U.S. and European universities,
Preismeyer, H. Richard and Kibak Batik 91989. "Discovering the Patterns including the University of California at Berkeley, the University
of Chaos." Planning Review 17 (November-December): 14-47. of Cincinnati, Temple University, and the Stockholm School of
Prigogine, Ilyaand Isabelle Stengers. 1984. Order Out of Chaos. London:
Economics. His research focuses on the dynamics of Marketing
Heinemann 9
systems, interfirm relations, and international marketing. His
Redmond, William H. 1991. "When Technologies Compete: The Role of
Externalities in Nonlinear Market Response." Journal of Product work has been published in journals such as the Journal of the
Innovation Management 8 (September): 170-183. Academy of Market Science, Journal of Macromarketing, Indus-
Rossiter, John R. and Larry Percy. 1987. Advertising and Promotion trial Marketing Management, Journal of Applied Psychology,
Management. New York: McGraw-Hill. and the European Journal of Marketing.