Regular Paper

A Computational Theory of Enterprise

Transformation
Zhongyuan Yu, William B. Rouse,* and Nicoleta Serban
Tennenbaum Institute, Georgia Institute of Technology, Atlanta, GA 30332-0210
A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

Received 13 August 2010; Revised 20 December 2010; Accepted 20 December 2010, after one or more revisions
Published online in Wiley Online Library (wileyonlinelibrary.com).
DOI 10.1002/sys.20188

ABSTRACT
This paper translates a qualitative theory of enterprise transformation into a quantitative, mathematical
theory. This enables computational exploration of the phenomena outlined in the theory, as well as the
sensitivity of these phenomena to a range of parameters intended to reflect the elements of the theory.
The overall question addressed is, What should an enterprise do in response to perceptions of impending
substantial change? A mixture of three types of responses is consideredpredict better, learn faster, and
act faster. As these responses all require investments, it is not the case that an enterprise should always
pursue all of them. Indeed, there are conditions under which an enterprise should not pursue any of them.
This paper elaborates the nature of these conditions, explains why they arise, and discusses the
implications for enterprises entertaining transformation. 2011 Wiley Periodicals, Inc. Syst Eng 14:
Key words: enterprise transformation; change strategies; computational modeling; change point
detection

1. INTRODUCTION

change. For example, the direction of the change may be clear,

but its timing and magnitude may be uncertain. The second
type of uncertainty concerns how best to respond to change.
What should ones offerings and processes become?
Given insights into or answers of these questions, how can
one best allocate resources between the enterprise one has and
the enterprise one is striving to become? Thus, the leaders of
these enterprises must wrestle with the question of whether
they should invest in becoming better at what they are currently doing versus investing in doing new things that will
better match emerging market desires. In other words, should
they focus on business process improvement or enterprise
transformation?
These are difficult issues. Gerstner [2002, p. 64] portrays
this difficulty, indicating Reengineering is like starting a fire
on your head and putting it out with a hammer. His approach
to enterprise transformation was to turn IBM into a marketdriven rather than internally focused, process-driven enterprise. A recent study of four successful
transformationsLockheed Martin, Newell Rubbermaid,
Reebok, and UPShas shown that there are a variety of

Contemporary enterprises face enormous uncertainties. Defense procurements may no longer be dominated by traditional weapon systems. Healthcare delivery may transition
from pay for services to payment for health outcomes. Intelligent sensing and control technology may morph the energy
ecosystem. There are many examples of substantial uncertainties in other industries such as education, finance, and food
[Rouse and Basole, 2010].
There are, at least, two types of uncertainty. First, the
enterprise may not be sure of the nature of the impending
*Author to whom all correspondence should be addressed (e-mail:
bill.rouse@ti.gatech.edu).
Contract grant sponsor: This research was supported, in part, by the Mark &
Kimberly Miller Charitable Foundation.
Systems Engineering
2011 Wiley Periodicals, Inc.

YU, ROUSE, AND SERBAN

approaches to transformation, but some common underlying

competencies including vision, leadership, strategy, planning,
communication, and cultural change [Rouse, 2011].
While current levels of uncertainty may seem unprecedented, this is not the case. We are just entering the 40th
decade of American experience, at least from the perspective
of European immigrants. In only seven of these decades has
there not been a war or economic crisis. The 1880s was the
most recent of such decades. Change and uncertainty are
strongly woven through American history.
In a study of 200 companies over the past 200 years in the
transportation, computer,1 and defense industries in the
United States, it was found that many attempted transformation in the face of changeand most failed [Rouse, 1996].
More recently, The Economist reported that the Fortune 500
has seen a 200% turnover in the past 25 years [Schumpeter,
2009]. Thus, the challenge of enterprise transformation has
been, and continues to be, very significant.
To provide a framework for exploration of the nature of
enterprise transformation, Rouse [2005, 2006] proposed a
theory of enterprise transformation. This theory focuses on
why and how transformation happens, as well as ways in
which transformation is addressed and pursued in terms of
work processes and the architecture of these processes. This
theory is stated as follows:
Enterprise transformation is driven by experienced and/or
anticipated value deficiencies that result in significantly redesigned and/or new work processes as determined by managements decision making abilities, limitations, and
inclinations, all in the context of the social networks of
management in particular and the enterprise in general.

The purpose of this paper is to translate this qualitative

statement of the theory into a quantitative, mathematical
theory to enable computational exploration of the phenomena
outlined in the theory, as well as the sensitivity of these
phenomena to a range of parameters intended to reflect the
elements of the theory. The overall goal is to address the
question, "What should an enterprise do in response to perceptions of impending substantial change? A mixture of
three types of responses is consideredpredict better, learn
faster, and act faster. As these responses all require investments, it is not the case that an enterprise should always
pursue all of them. In fact, there are conditions under which
an enterprise should not pursue any of them. This paper
elaborates the nature of these conditions, explains why they
arise, and discusses the implications for enterprises entertaining transformation.

2. CONCEPTUAL FORMULATION
2.1. Model Constructs
The central constructs of the theory are value, work processes,
decision making, and social networks. Value is a measure of
1
The roots of the computer industry in the second half of the 20th century
can be traced back to the cash register and typewriter industries in the 19th
century. Indeed, the leaders in these two industries were among the key
players in the early computer industry.

Systems Engineering DOI 10.1002/sys

the extent to which an enterprise provides a market what the

consumers in this market want. Increasing variations of offerings from what consumers want results in decreasing value.
We hasten to acknowledge that value can also accrue from
providing customers an offering they did not expect, e.g.,
Apples iPhone. This leads to innovation whereby the market
changes its desires. Market innovators are usually not transformers; their innovations cause the other players in the
market to have to transform. Thus, for example, Wal-Mart
innovated in the retail marketplace; K-Mart and Sears were
thereby forced to transform.
Work processes are the means for translating an enterprises intentions into market offerings. These processes depend on people, information, facilities, and equipment.
Changing work processes requires investments, often substantial investments.
There are several aspects of decision making. One concerns what is happening in the market. Is the value of offerings
increasing or decreasing or holding steady? Another decision
concerns how best to adapt to needed changes. Should one
invest in predicting better, learning faster, or acting faster?
As elaborated later, another possibility is doing nothing
different. This has the advantage of being easy to do and
perhaps not requiring any additional investment. Indeed, we
have found in working with numerous large enterprises that
the status quo is often a very compelling alternative. We later
discuss the circumstances under which the status quo is the
best choice.
Social networks play several roles. First, of course, they
constitute how work gets done in terms of who does what,
who knows what, who knows who, and so on. They can also
facilitate or impede change. Social networks can facilitate
change by embracing and rapidly diffusing change as has been
seen recently for e-mail and smart phones.
Social networks can also impede change, acting in ways
analogous to an enterprise immune system [Rouse, 1998].
One way to impede change is to deny that it is needed. This
delays recognition of market signals that the value of an
enterprises offerings is declining. Another impedance is reluctance to change work processes. Competencies that provided competitive advantage in the past may be preserved,
perhaps receiving sustained investment, despite their decreased relevance to competing in the future.

2.2. Conceptual Model

Figure 1 provides a high level conceptual model of the theory
of enterprise transformation that forms the basis for the
mathematical model introduced in the next section. This
mathematical model makes the concepts in Figure 1 quite
precise.
The key elements of this model include the following:
Management: Represents decisions to allocate resources to remediate value deficiencies and maximize
expected utility
Production: Represents mapping of resources, via labor, to products and services over time

A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

which contributed to Senges more recent view of organizational learning [Senge, 1990].
Markides [1999] addressed the dynamic nature of strategy
in terms of continuous reassessment and reformation. Miles
and colleagues [1978] reviewed how organizations define
strategies relative to product-market domains and then align
structure and processes to pursue these strategies. Moncrieff
[1999] contrasted planned and emergent strategy in terms of
five elements:
Figure 1. Elements of the theory.

Market: Represents mapping from products and services to value that leads to revenues, profits, and cash
flows over time
Social Network: Represents allocation of human attention to deploy resources, including provision of information for decision making.
The overarching question to be asked of this model concerns how market uncertainties and social network characteristics affect the decisions management must make to
allocate resources, including how the information available,
plus predictions, affects the possibility of making well-informed decisions.

2.3. Broader Theories

Before addressing the mathematical formulation of the conceptual model in Figure 1, it is important to discuss the extent
to which this model aligns with contemporary management
theory. In other words, how does the theory of enterprise
transformation underlying this model fit in with broader
theories of business strategy?
It has long been recognized that strategy formation is not
a static mapping, nor a simple mapping, from industry structure to strategic plans to exploit this structure to maximize
profits [Porter, 1991]. Even earlier, Schendel and Patton
[1978] outlined the multidimensional nature of strategy,
where profit is but one element. Mintzberg [1978] elaborated
the interaction of the organization with the environment,

Organizational intentions
Organizational response to the environment
Dynamics of the organization
Alignment of action with intentions
Strategic learning.

These elements are reflected in Figure 1, albeit with more

contemporary terminology.
There are additional important subtleties. Eisenhardt and
Martin [2000] elaborated a resource-based view of strategy
and contrasted moderately versus highly dynamic markets.
They summarized evidence that analytical strategies work
best for the former, and experimental approaches work better
for the latter. However, such differences are not inherent.
Kelly and Amburgey [1991] reported that organizational
change is not always related to discontinuous environmental
changes, nor does it always relate to chances for survival.
Thus, it would be folly to argue that the theory presented
in this paper reflects how companies will inherently respond
to circumstances that should prompt enterprise transformation. Instead, the goal is to elaborate the strategic choices
available to companies who address transformation in this
manner and the conditions under which these choices make
sense.

3. MATHEMATICAL FORMULATION
The goal is to convert the above conceptual model to as simple
a mathematical model as possible that will enable representation of the central phenomena of interest and support
computational exploration of the nature of these phenomena,
including their sensitivity to key parameters. Although our
model is simplified given the complexity of enterprise trans-

Table I. Summary of Key Parameters

Systems Engineering DOI 10.1002/sys

YU, ROUSE, AND SERBAN

formation, it enables fundamental understanding of how key

characteristics outlined in Figure 1 impact transformation at
a higher level.
The model elaborated in this section has four key parameters as summarized in Table I. These parameters reflect central
elements of the theory of enterprise transformation. The next
subsection presents a series of simulation experiments aimed
at illustrating the sensitivity of market performance to the
model parameters and the role of these parameters in investment strategies and in driving fundamental changes faced by
an enterprise.

3.1. Market and Value Models

The market defines value. Hence, elaboration of the model
begins with the market. The evolution of the markets desires
is described by
XM(t + 1) = XM(t) + (1 ) W(t).

(1)

The markets desire at time t + 1, XM(t + 1), is driven by its

desire at time t, XM(t), and a stochastic process denoted by
W(t). While W(t) could take many forms, in this paper it is
assumed to be a Gaussian variable with mean of 100 and
variance of 10. According to this model, the smaller alpha is,
the more weight placed on the randomness of the market
implying an instable market.

The markets desire, XM, should be viewed as an abstract

set of product or service characteristics sought by the market,
e.g., price, performance, fuel economy, etc. Changing market
desires requires a company to change their offerings accordingly if these offerings are to continue to be desired by the
market. Value is defined as a function, F, of the difference
between market desires, XM, and the characteristics of a
companys offering, XO, as denoted in
VM(t) = F[XM(t), XO(t)].

This function could take on various forms. The functional

form used for the simulation experiments discussed in this
paper is given by
VM(t) = EXP{ ABS[XM(t) XO(t)]}.

(3)

Thus, market value VM is assumed to decrease exponentially

with absolute differences between market desires and market
offerings. Markets with higher lambda are more discriminating than markets with lower lambda.
Figure 2 depicts the impact of alpha on market desires
(upper plot) and perceived market value (lower plot). When
alpha is large (e.g., alpha 0.8), the market is more stable,
resulting in more predictable market desires and market value.
When alpha is smaller (e.g., alpha < 0.5), the market becomes

Figure 2. Impact of market dynamics.

Systems Engineering DOI 10.1002/sys

(2)

A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

noisy, resulting in a more random market behavior regardless

of whether there is a significant trend or not. Later results will
show that as alpha decreases, a companys abilities to change
successfully greatly diminish. This has a direct impact on the
choice of the investment strategy.
Figure 3 shows the impact of lambda, representing the
extent to which customers are highly discriminating relative
to differences between XO and XM. Small lambda indicates
forgiving customers, while larger lambda indicates discriminating ones. As expected, perceived market value decreases
substantially when lambda increases.
An important issue for a company concerns how quickly
they learn about market value, especially when the nature of
value is changing [Eisenhardt and Bourgeois, 1989]. Unfortunately, many companies are sometimes slow to recognize
that they are no longer delivering value as in the past. This is
due in part to the practicalities of data collection and interpretation. However, there are social factors that affect how
quickly a company learns, as elaborated below.
The delay in recognizing value changes is represented by
differentiating perceived value, VP, from actual market value,
VM, as defined by
VP(t + 1) = 1 VM(t) + 2 VM(t 1) + 3 VM(t 2) + .
(4)
Ideally, 1 equals 1 and the higher order deltas 2, 3, 4, ...
are zero. Unfortunately, the higher order deltas may dominate
what a company perceives; e.g., what used to be highly valued
may be seen by the company as still highly valued, although
it may not be, as recently illustrated by the automobile companies. In this paper, to avoid confusion and extensive notation, when we say that delta equals 2, we mean that the second
order delta 2 equals 1 while the other deltas are all zero, for
example.
The challenge for a company in a changing market is to
get XO as close as possible to XM to yield the highest VM,
subject to the difficulty of only knowing VP. Closeness is
most important for highly discriminating markets, although

large variations between XO and XM will be strongly penalized by all markets.

3.2. Work Process Model

The work processes of a company determine the characteristics of the companys product and service offerings. If XM
changes suddenly, perhaps due to an economic crisis, a company may find it difficult to immediately shift XO to match
XM. This difficulty is represented by
XO (t + 1) = XO (t) + (1 ) XF (t + T)

(5)

A companys offerings at time t + 1, XO(t + 1), are highly

influenced by its offerings at time t, XO(t), as well as its
forecast, XF(t + T), of future market desires, XM. Offering
design, equipment, and facilities may be such that they cannot
be morphed quickly; e.g., shifting from producing pickup
trucks to economy cars may take time. Beyond being a
practical issue, such changes may be impeded by social
forces, as discussed below.
Figure 4 illustrates the impact of the beta parameter on the
market offering XO (upper plot) and on the perceived market
value VP (lower plot). For example, in the upper plot, at time
t = 120, the market desire is around 220. However when beta
equals 0.8, the company can only respond with an offering of
about XO = 180, which will result in lower market value. Thus
lower values of beta imply that the company can act more
quickly and is better in tracking changes in market desires.
On the other hand, the sluggishness of the company has a
substantial impact on its responsiveness to change.
As indicated in Eq. (5), the characteristics of a companys
offerings are affected by previous offerings XO(t), as well as
the companys market forecasts, denoted by XF(t + T) with
prediction horizon T in Eq. (6).
^
^
^
___(t), XM
___(t 1), XM
___(t 2), ]
XF(t + T) = F[XM

(6)

Figure 3. Impact of market discounting.

Systems Engineering DOI 10.1002/sys

YU, ROUSE, AND SERBAN

Figure 4. Impact of enterprise social network.

The many nuances and subtleties of Eqs. (1)(6), including

the impacts of , , , and , are discussed in the subsequent
section describing the simulation experiments.

3.3. Decision Making

In this section, attention shifts to the nature of management
decision making in the context of the proposed computational
theory of enterprise transformation. Succinctly, what choices
does management have and how should resources be allocated
among these alternatives? There are four major types of
decisions to be addressed. First of all, given a sluggish company with delayed market feedback, management could decide to invest in making better and longer-range predictions,
i.e., predict better. This is a good idea to the extent that alpha
is large enough to enable accurate predictions. For example,
this works really well for sales that are driven by demographics, e.g., it is easy to predict how many young children will
reach school age next year.
Another, complementary, decision would be to learn faster.
This would involve investing in the infrastructure and people
for information gathering and interpretation in hope of shifting the weights in Eq. (4) towards more recent market data.
There can be, as noted earlier, social forces that can inhibit
this. To counteract such forces, companies often employ focus
groups of consumers to get external views of the value of their

Systems Engineering DOI 10.1002/sys

products and services. Later discussions consider the impact

of such investments on V/ , which should inform investment decision making.
A third complementary decision would be to act faster.
This would involve investing in people, processes, equipment,
and so on to decrease beta in Eq. (5). The result would be a
less sluggish company that can change XO much faster in
response to forecasts of changing XM. A good example of this
is Hondas investment in flexible production lines that enabled shifting production to the smaller cars in its lineup when
demand for larger cars faded. The impact of such investments
on V/ is also discussed later.
Finally, management could focus on more quickly detecting market changes, i.e., detect faster. This involves the topic
of change point detection, which is elaborated in the discussion of the simulation experiments. Put simply, there is a
tradeoff between the power of change point decisions, e.g.,
the probability of correctly rejecting the null hypothesis of no
change, and the time it takes to conclude that a change has
happened. The power threshold chosen can be highly influenced by social forces and cultural attributes of an enterprise.

3.4. Social Networks

A company is not just a set of human, physical, and financial
resources organized around creating and delivering what the

A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

market desires. A company is also a social system that tends

to have enormouspositive and negativeimpact on the
enterprises abilities to change [Rouse, 1993, 2005, 2006,
2007 and the references therein]. This social system can
facilitate change via rapid communication and support of
appealing changes. It can also inhibit change, much like an
immune system, when the changes being entertained conflict
with the concerns, values, and perceptions of key stakeholders.
Beta and delta within the model elaborated in the previous
section reflect social phenomena just as much as they reflect
practical limitations on change. Acting faster, by decreasing
beta to enable doing new things, will usually compete for
resources with the status quo. Indeed, in times of stressful
change, the as is enterprise can consume enormous resources trying to sustain the incumbent business model. Jobs,
livelihoods, and political support may hang in the balance. For
example, shifting investments from bending metal to coding
software will likely encounter stiff resistance from those
weaned on bending metal.
Learning faster, by shifting delta to gain market knowledge
more quickly, will compete with organizational belief systems about market value and the extent to which the status
quo remains competitive [Rouse, 1993, 1998]. Denial of
change is often easier to argue than acceptance of change
coupled with a lack of action. The insularity typical of large
successful enterprises is likely one of the reasons the Fortune
500 experiences such high levels of turnover [Schumpeter,
2009].
There is also a social aspect to the statistical power required for agreement that change has happened. If the threshold is very high, say 9095%, then the company will be quite
delayed in reacting to change. On the other hand, if the
threshold is quite low, perhaps 5060%, then the company
will react to apparent changes that have not actually occurred.
There are typically very strong social forces against abandoning a business model that was the source of previous success.
This perspective can be quite reasonable when it is not at all
clear what new business models will succeed.

4. SIMULATION EXPERIMENTS
To support the qualitative insights into the enterprise transformation theory, computational experiments were conducted to
assess the sensitivity of a companys performance (i.e., average market value realized) to market trends and model parameter settings (i.e., , , , and ). Since the theory of
transformation is driven by recognition of experienced or
anticipated value deficiencies, these experiments required a
mechanism for a company to detect change. Consequently,
this section begins with consideration of change point detection. With this mechanism defined, focus then shifts to the
three strategic decisions that a company should entertain
predicting better, learning faster, and acting faster.

4.1. Change Point Detection

Change Point Detection Theory. The observations in this
study, Yj, j = 1, . . . , n, are the perceived market values VP

realized at different time points tj, j = 1, . . . , n. A company

is interested in deciding whether there is a change in VP within
a given period of time.
A simple linear model for the perceived market values is
given by
Yj = b0 + b1 tj + j for j = 1, . . . , n,

(7)

where the js are error terms and assumed to be identically

distributed but correlated. Under this model, the hypothesis
of a change point in the perceived market value is
H0: Yj = b0 + b1 tjj for j = 1, . . . , n,

(8)

vs.
HA: Yj = b0 + b1 tj if j < and
Yj = b0 + b1 tj if j

(9)

where b0 b0 or/and b1 b1. That is, under the alternative

hypothesis the regression line changes its slope and/or its
intercept, which is an indication of a change point at tj in the
observed data. For detecting a change point, one computes a
test statistic Uj at a given time point tj using the approach
elaborated in the Appendix. High U values indicate a potential
change in intercept and/or slope at the corresponding time
points.
However, a high U value alone does not guarantee a change
point; one has to make an inference of how high the U value
needs to be using the test p-value. In this study, this is achieved
by using bootstrap sampling, since the distribution of the test
statistic under the null hypothesis does not have a closed form
expression. This is because the observations Yj, j = 1, . . . , n,
are serially correlated, and, therefore, the classical assumption
of independence does not hold. Application of the bootstrap
sampling procedure for obtaining the significance level and
the p-value is described in the Appendix. If the p-value is
lower than a preset test significance level, specifically 0.05 in
this paper, one accepts the alternative hypothesis of a change
point. In the context of this study, the p-value corresponds to
the probability of deciding a change in the market value at a
given time point although no change has happened. The
smaller the p-value, the smaller is the probability of making
an error of deciding that there is a change in the perceived
market value when there is in fact not a change.
An important aspect of the testing procedure discussed
above is the power of detecting a change in the market value
given market trends and parameter settings (i.e., , , , and
). For this, Monte Carlo simulations are employed as discussed in the Appendix. The power of the test provides a
probabilistic assessment of the effectiveness of the change
point detection procedure described in this paper. The higher
the power, the more accurate the decisions are.
The change point detection procedure is applied to selected
observed time pointsin this study, the actual change point
is time point 101 and the selected observed time units are from
101 to 110. Therefore, for each selected time point, we derive
their corresponding U values, U101, U102, . . . , U110 and

Systems Engineering DOI 10.1002/sys

YU, ROUSE, AND SERBAN

p-values, p101, p102, . . . , p110. Then for the same parameter

settings, the above simulation procedure is repeated 60 times.
The power of the detection test at each of the selected time
points is obtained by counting the number of simulations that
have p-values less than the 0.05 preset significance level.
Detection time is defined as the smallest number of time
points needed to detect a change. That is, the detection time
is the first time point among t101, . . . , t110 for which the
corresponding power of the detection test is larger than a
companys decision threshold.
The length of time to discover a change in the market value
varies from one company to another and from one business
setting to another, as reflected by market trends and model
parameters (i.e., , , , and ). A company may choose a
different decision threshold, i.e., power threshold, to detect a
change. For example, an aggressive company may set the
decision threshold as 60%, implying the company will take
action when it is 60% sure that there is a change, whereas a
conservative company may set this threshold as high as 95%.
Change Point Detection Results show how trend, alpha,
beta, lambda, and delta, as well as the companys decisionmaking criteria, affect the delay in discovering that change
has happened in the market according to the simulation.
Effect of trend: Not surprisingly it is easier for companies
to detect sudden and rapid changes than slow and stable
changes.
Effect of market dynamics (): Stable markets (higher )
naturally will display higher power to detect a change point
in the market value. However, when is quite high ( 0.5),
the sluggishness reduces the effect of alpha; therefore, the
differences in market value trends for varying alpha are not
significant.
Effect of enterprise social network (): Sluggish companies (higher ) can detect change more easily, although they
cannot act sufficiently quickly to compensate for this change.

In this way, a companys sluggishness can act as a filter on

market signals, which makes change more pronounced.
Effect of market social network (): Increasing delta will
directly delay detection time. The increased delay of detection
time caused by the delta is approximately the same as the
observation delay of the market value due to larger . Assume
the actual change happens at period 101, when there is no
delay in learning the market value (1 = 1), it is quite possible
to detect the change at period 102. When there is 2-time-unit
delay (3 = 1), one will likely detect change at period 104.
Effect of market discounting (): Discriminating customers (higher ) are those that sharply discount the value of
products or services that do not closely match their desires.
Such discounting will delay the detection of change, because
the volatility of market value caused by crisply defined market
desires rather than trend changes impedes the detection of
change.

4.2. Decision Making

Predict Better. In a stable market, given a sluggish company
with delayed market feedback, management could decide to
invest in better and longer-range predictions to force the
company to change by using predictions of future market
desires to drive the current market offering, much in the same
way that a supertanker captain uses the longer term desired
path to drive shorter-term steering decisions. This is accomplished by setting XO(t + 1) = XO(t) + (1 ) XF(t + T)
with prediction horizon T (T time units ahead), where T can
be determined by beta, the sluggishness of a company.
Figure 5 shows how different look-ahead periods can affect
market value. If the enterprise social network () is equal to
0.5, it is better to look ahead 2 time units, as the dotted curve
has the highest market value. For each beta, there are always
one or two best look-ahead periods. Figure 6 shows how
look-ahead time units should be adjusted for beta, and how
different look-ahead time units affect average market value.

Figure 5. Market value vs. number of time units look ahead.

Systems Engineering DOI 10.1002/sys

A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

When beta equals 0.5, it is better to look ahead 2 time units,

as we can see the dashed curve has the highest average market
value. The trend and alpha do not change where the maxima
occur, but change how pronounced the maxima are.
Learn Faster. The delay of change point detection time
leads to lost value. The lost value is calculated by the timecumulative value until the company catches up with the target
value that would have occurred had detection not been delayed (the area between the nondelayed and delayed detection
time). Results show that lower and higher , i.e., sluggish
companies in unstable markets, experience greater loss when
there is a delay in detection of a change. The market social
network () will directly affect the detection time. The increased delay of detection time caused by delta is approximately the same as the delta increase.
Figure 7 shows V/, which is defined as the average
difference of total lost value caused by the 1 unit detection
time change as a function of the market dynamics (). From
this figure we can see that higher results in shorter detection
times and smaller total lost value. Also, the longer companies
wait to make decisions, the more loss per time unit they will
encounter.
Act Faster. One reason for low market value is the sluggishness of a company (reflected by high beta). A decision to
act faster (decrease beta) involves investing in people, processes, equipment, and so on. An enterprise is interested in the

impacts of such investments on V/. These results are

sufficiently compelling to discuss them in the next section.

4.3. Summary of Results

The simulation results discussed in this previous section can
be summarized as follows: Change point detection time decreases with increasing trend and alpha, and increases with
increasing lambda and delta. The best case is a strong trend
in a stable market for a responsive company that learns
quickly and has forgiving customers. The worst case is a mild
trend in a noisy market with a sluggish company that is slow
to learn and has highly discriminating customers. Interestingly, sluggishness is good in the sense that it decreases
detection time but, on balance, is bad in that it inhibits
responding to what is detected.
Investments in predicting better can compensate for high
beta, with the optimal prediction period increasing with beta,
independent of alpha and trend. Thus, if ones company is
more like a supertanker than a speedboat, longer-term predictions should drive decisions regarding how best to track
market desires. Noisier markets will result in less accurate
predictions, but nevertheless these predictions should drive
decisions.
Learning faster results in less lost value due to delayed
detection of change points, which provides greater returns

Figure 6. Average market value vs. beta.

Systems Engineering DOI 10.1002/sys

10

YU, ROUSE, AND SERBAN

Figure 7. Investments in learning faster, V/.

(less lost value) for companies with low beta in markets with
high alpha. Thus, learning faster helps most when the market
is more predictable and a company can respond quickly.
Acting faster by investing in decreasing beta yields improved market value, independent of alpha for low initial
market value, but very much dependent on alpha for higher
initial market value. An enterprise can only approach maximum value in very predictable markets when it can act
quickly.

5. IMPLICATIONS FOR INVESTMENTS

These results beg the question of how best to allocate scarce
resources to improve market value. Acting faster is powerful,
particularly in stable markets with strong trends. If the market
is very noisy or the trend is weak, then investing to reduce
beta may not be worthwhile. In such situations, it may be best
to stay with the status quoafter all, this may not require any
additional investment and the company already knows how to
do it. Changing, in contrast, will require investment and, if the
market is sufficiently noisy, may result in a situation just as
bad (or worse) as not changing at all.
The overarching question concerns: When to act and when
not to act? We answer this question by assessing the return
V/ for varying values of alpha, beta, and lambda. In
general, V/ should exceed some threshold to warrant

Systems Engineering DOI 10.1002/sys

investing in change. The threshold will depend on how V

and are monetized, i.e., what V is worth and what
costs. One could easily imagine, for example that V/
>1 could be the investment criteria.
Figure 8 shows how much market value increases for beta
decreasing from 0.1 to 0. Under this scenario, the return is
high when alpha and lambda are both high. In other words, if
the market is more predictable, customers are discriminating
and one can act quickly, the return is high. On the other hand,
if the market is noisy or customers are not discriminating, the
likely return (V) from investing is not worth the cost of the
investment ().
Figure 9 illustrates how much market value increases for
beta decreasing from 0.5 to 0.4. Under this scenario, the return
is high for median alpha and median lambda. This may seem
counterintuitive. The reason is that there are two underlying
phenomena.
When customers are forgiving, it is not worth investing
as one is already hitting the target of maximum value.
When customers are highly discriminating, it is not
worth investing because one is very unlikely to hit the
target.
When the market is less noisy, the investment is worth
more until the market becomes highly predictable because in this case one would already be on target.

A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

11

Figure 8. Investments in acting faster, V/ ( = 0.00.1).

Figure 10 illustrates how much market value increases for

beta decreasing from 0.9 to 0.8. If customers are discriminating (high ) and one acts slowly (high ), one is unlikely to
hit the target. On the other hand, if the customers are not
discriminating and the market is noisy, it may be worth
investing to act a bit faster as it increases the chances of hitting
the target, considering how poorly one is currently performing.

As indicated earlier, predicting better and learning faster

can always be justified, with the prediction horizon adapted
to the level of sluggishness of the company. The returns on
such investments will be greater for more predictable markets
and less sluggish companies. Considering the likelihood that
such investments will be much smaller than those needed to
decrease beta, these investments are much easier to justify.

Figure 9. Investments in acting faster, V/ ( = 0.40.5).

Systems Engineering DOI 10.1002/sys

12

YU, ROUSE, AND SERBAN

Figure 10. Investments in acting faster, V/ ( = 0.80.9).

To summarize, investing in transformation is likely to be

attractive when one is currently underperforming and the
circumstances are such that investments will likely improve
enterprise performance. In contrast, if one is already performing well, investments in change will be difficult to justify.
Similarly, if performance cannot be improveddue to noisy
markets and/or highly discriminating customersthen investments may not be warranted despite current underperformance.

6. CONCLUSIONS
The computational theory elaborated in this paper predicts
that companies will transform their enterprise by some combination of predicting better, learning faster, and acting faster,
as long as the market is sufficiently predictable to reasonably
expect that transformation will improve the market value the
company can provide. If this expectation is unreasonable, then
companies will sit tight and preserve resources until the
market becomes more fathomable.2
The theory elaborated in this paper is premised on the
notion that companies make transformation decisions in response to the dynamic situations in which they find themselves. These decisions are affected by both what the company
knows (or perceives) and the companys abilities to predict,
learn, and act. Indeed, decisions to transform abilities to
predict, learn, and act reflect desires to fundamentally change
the companys overall ability to create market value. In this
way, transformation decisions can enhance a companys abilities to address the ongoing and anticipated transformations
needed for success in dynamic markets.

APPENDIX: TEST FOR A CHANGE-POINT IN

LINEAR REGRESSION
The observations are (Xj, Yj), j = 1, . . . , n, where X jg = (1,
xj,1, . . . , xj,p are a set of predictors. In this article, the response
Yj corresponds to perceived market value whereas the predictor is time (p = 1).
The null hypothesis of interest is that the observations
satisfy a linear regression model:
H0: Yj + Xjg + j for j = 1, . . . , n.

(10)

The alternative hypothesis is HA, that there is a change-point

such that
Yj = Xjg + j if j <

and

Yj = Xjg + j if j > for .

(11)

Here j for j = 1, . . . , n are errors assumed to be identically

distributed but correlated.
In this study, the likelihood ratio test provided by Kim
[1994] was applied to detect changes in linear regression
models with the modification of the p-value.
The likelihood test statistic is defined as follows: For
k = 1, , p, let ak be the (p + k) 1 vector with 1 and 1 at
the (2k 1)st and the (2k)th component, respectively, and 0 in
all other components, and let Xk,j be the n (p + k) matrix such
that
Xk,jek,i = (1jg, 0jg) for i = 1,

(12)

Xk,jek,i = (0jg, 1jg) for i = 2,

(13)

This can be observed during the current (2010) economic conditions where
companies are making strong profits, but hoarding cash rather than investing
in increased capacity and/or new offerings due to great uncertainties about
where the economy is headed.

Systems Engineering DOI 10.1002/sys

A COMPUTATIONAL THEORY OF ENTERPRISE TRANSFORMATION

g
Xk,jek,i = x i1 ,1, . . . , x i=1 ,j 0nj
for i = 3, 5, . . . 2k1,(14)
2

2
g
Xk,jek,i = 0nj
, x i1 ,j+1, . . . , x i1 ,n for i = 4, 6, . . . 2k,(15)

2
2

Xk,jek,i = x i1 ,1, . . . , x i1 ,n for i = 2k + 1, 2k + 3, . . . (, 16)

2
2
Xk,jek,i = x i ,1, . . . , x i ,n for i = 2k + 2, 2k + 4, . . . , (17)
2
2
where 1j is the length vector of 1s, 0j is the j length vector of
0s, and ek,i is the (p + k) length vector with 1 in the ith
component and with 0 in all other components. For
k = 1, , p, we further define
g
g
a k,g j(X k,j
X k,j)1Xk,j
Y
,
Uk(j) =
1

g
g 2
g
a k,j
(X k,j
X k,j)1ak,j

(18)

where Y = (Y1, . . . , Yn). Lastly, the test statistic referred as

the U value in this paper is
^
^
U(Y) = 2 max ||U(j)||2 = 2 max {U21(j) + + U2p(j)}
p j np
p j np
(19)
^ 2 is the maximum likelihood estimate of 2 under
where
H0.
Worsley [1983] studied the likelihood ratio test (LRT) for
testing H0 against HA, and provided an upper bound for the
significance level of the test. However, in this case study, the
assumption of independence among Y1, . . . , Yn does not hold
as they are serially correlated, and therefore, existing LRT
results do not apply. Instead a resampling method called
parametric bootstrap is used to sample from the null distribution of the test statistic defined above which will further
provide an approximation of the p-value of the change point
hypothesis test. Specifically, a sample of the test statistic is
obtained under the null hypothesis as follows:
1. Fit a linear regression under the null hypothesis:
^
^ where
^ are estimated coefficients. Extract
H0: Yj = Xjg
^ , j = 1, . . . , n.
the residuals Rj = Yj Xjg
2. Permute the residuals R1, . . . , Rn to obtain a new sample
from the error distribution: R1 , . . . , Rn.
3. Obtain a new sample from the null hypothesis:
^ + R, j = 1, . . . , n.
Yj = Xjg
j
4. Repeat 23 for B (= 1000) times.
Using this re-sampling technique, we obtain B samples
b
from the null hypothesis Yb = (Yb
j , . . . , Yj ) for b = 1, . . . ,
B and for each sample we can compute the U value,
Ub = U(Yb). We approximate the p-value using
B

1
p-value =
B

b=1

I(U(Y) U(Y b)).

(20)

13

An important evaluation criterion of a hypothesis test is its

power, defined as the probability of rejecting the null hypothesis when the alternative hypothesis is true. In the context of
this problem, the power of the test is translated as the probability of detecting a change when a company indeed has a
change in the market value. Similarly to the computation of
the p-value, because one does not have a close form expression for the distribution of the test statistic, Monte Carlo
simulations are employed to obtain an approximation of the
power of the change point detection test. Specifically, given
the parameter settings (i.e., trend, , , , and ), S = 60
samples are generated from the model. One then evaluates the
number of times the p-value is smaller than a significance
level (0.05 in this paper) divided by the number of samples,
S.
Remark: In change point detection theory, one tuning
parameter that may affect the reliability measured by the
p-value and the accuracy measured by power of the test
hypothesis for detecting a change is the number of observations set before (past) and after (future) a test time point. The
study shows that the test results are not sensitive to the
selection of the future and past points as long as the ratio
between them is less than 2/3 and as soon as the number of
past time points is between 20 and 40 and the number of future
time points is 515, respectively. However, when, for example, the number of past time points is too large or the number
of future time points is too small, the accuracy of the decision
regarding change in the market decreases.

REFERENCES
K.M. Eisenhardt and L.J. Bourgeois, Making fast strategic decisions
in high-velocity environments, Acad Management J 32(3)
(1989), 543576.
K.M. Eisenhardt and J.A. Martin, Dynamic capabilities: What are
they? Strategic Management J 21(1011) (2000), 11051121.
L.V. Gerstner, Jr., Who says elephants cant dance? Inside IBMs
historic turnaround, Collins, New York, 2002, 64.
E.S. Hanawalt and W.B. Rouse, Car wars: Factors underlying the
success or failure of new car programs, Syst Eng 13(4) (2010),
189403.
D. Kelly and T.L. Amburgey, Organizational inertia and momentum:
A dynamic model of strategic change, Acad Management J 34(3)
(1991), 591612.
H.-J. Kim, Tests for a chance-point in linear regression, IMS Lecture
Notes, Monograph Series, Volume 23, Institute of Mathematical
Statistics, Beachwood, OH, 1994.
C.C. Markides, A dynamic view of strategy, Sloan Management Rev
40(1) (1999), 5563.
R.E. Miles, C.C. Snow, A.D. Meyer, and H.J. Coleman, Jr., Organizational strategy, structure, and process. Acad Management Rev
3(3) (1978), 546562.
H. Mintzberg, Patterns in strategy formation, Management Sci 24(9)
(1978), 934948.
J. Moncrieff, Is strategy making a difference? Long Range Plan Rev
32(2) (1999), 273276.
M.E. Porter, Towards a dynamic theory of strategy, Strategic Management J 12(S2) (1991), 95117.

Systems Engineering DOI 10.1002/sys

14

YU, ROUSE, AND SERBAN

W.B. Rouse, Catalysts for change: Concepts and principles for

enabling innovation, Wiley, New York, 1993.
W.B. Rouse, Start where you are: Matching your strategy to your
marketplace, Jossey-Bass, San Francisco, 1996.
W.B. Rouse, Dont jump to solutions: Thirteen delusions that undermine strategic thinking, Jossey-Bass, San Francisco, 1998.
W.B. Rouse, A theory of enterprise transformation, Syst Eng 8(4)
(2005), 279295.
W.B. Rouse (Editor), Enterprise transformation: Understanding and
enabling fundamental change, Wiley, Hoboken, NJ, 2006.
W.B. Rouse, People and organizations: Explorations of human centered design, Wiley, Hoboken, NJ, 2007.

W.B. Rouse, Necessary competencies for transforming an enterprise, J Enterprise Transformation 1(1) (2011), in press.
W.B. Rouse and R.C. Basole, Understanding complex product and
service delivery systems, Handbook of service science, P.
Maglio (Editor), Springer, London, 2010, pp. 461480.
D. Schendel and G.R. Patton, A simultaneous model of corporate
strategy, Management Sci 24(15) (1978), 16111621.
J. Schumpeter, Taking flight, The Economist (September 19, 2009),
78.
P.M. Senge, The fifth discipline: The art and practice of the learning
organization, Doubleday/Currency, New York, 1990.
K.J. Worsley, Testing for a two-phase multiple regression, Technometrics 25 (1983), 3542.

Zhongyuan (Annie) Yu is a Ph.D. student in the School of Industrial and Systems Engineering of Georgia Tech and a
Graduate Research Assistant at the Tennenbaum Institute. Her research focuses on economic decision analysis and
human decision making. Zhongyuan has various intern and project experiences in fields ranging from manufacturing
to supply chain management, to airline operations research, to real estate consulting and banking, and she has published
related papers, such as The Application of Industrial Engineering in Manufacturing Management and Modeling and
Solving the Spatial Block Scheduling Problem in Shipbuilding Industry using Particle Swarm Optimization.
Zhongyuan received a B.S. in Mechanical and Industrial Engineering in Tongji University (Shanghai, China) with a
minor in Journalism at Fudan University (Shanghai, China), and an M.S. in Industrial Engineering at Georgia Tech.

Bill Rouse is the Executive Director of the Tennenbaum Institute at the Georgia Institute of Technology. He is also a
professor in the College of Computing and School of Industrial and Systems Engineering. His research focuses on
understanding and managing complex public-private systems such as healthcare, energy and defense, with emphasis
on mathematical and computational modeling of these systems for the purpose of policy design and analysis. Rouse
has written hundreds of articles and book chapters, and has authored many books, including most recently Economic
Systems Analysis and Assessment (Wiley, 2011), People and Organizations: Explorations of Human-Centered Design
(Wiley, 2007), Essential Challenges of Strategic Management (Wiley, 2001), and the award-winning Dont Jump to
Solutions (Jossey-Bass, 1998). He has edited or co-edited numerous books including Engineering the System of
Healthcare Delivery (IOS Press, 2010), The Economics of Human Systems Integration (Wiley, 2010), Enterprise
Transformation: Understanding and Enabling Fundamental Change (Wiley, 2006), Organizational Simulation: From
Modeling & Simulation to Games & Entertainment (Wiley, 2005), the best-selling Handbook of Systems Engineering
and Management (Wiley, 1999, 2009), and the eight-volume series Human/Technology Interaction in Complex Systems
(Elsevier). Among many advisory roles, he has served as Chair of the Committee on Human Factors of the National
Research Council, a member of the U.S. Air Force Scientific Advisory Board, and a member of the DoD Senior Advisory
Group on Modeling and Simulation. Rouse is a member of the National Academy of Engineering and has been elected
a fellow of four professional societiesInstitute of Electrical and Electronics Engineers (IEEE), the International
Council on Systems Engineering (INCOSE), the Institute for Operations Research and Management Science (INFORMS), and the Human Factors and Ergonomics Society (HFES).

Nicoleta Serban is an assistant professor in the School of Industrial and Systems Engineering at Georgia Institute of
Technology. She received her Ph.D. and M.S. in Statistics from Carnegie Mellon University. She also holds a B.S. in
Mathematics and an M.S. in Stochastic Processes and Theoretical Statistics from the University of Bucharest. Dr.
Serbans research crosses multiple disciplines including methodological statistics, molecular biology, healthcare,
industrial engineering and socio-economics. Her primary methodological contributions in statistical research are for
the analysis of multiple time-varying random functions. Dr. Serbans research has been published or accepted in more
than 15 journal articles, most in top journals in statistics, engineering and biology.

Systems Engineering DOI 10.1002/sys