Principles of Strategic Data

Science
Creating Value from Data, Big and
Small

Peter Prevos
This book is for sale at http://leanpub.com/strategic_data_science

This version was published on 2019-03-08

This is a Leanpub book. Leanpub empowers authors and

publishers with the Lean Publishing process. Lean Publishing is
the act of publishing an in-progress ebook using lightweight tools
and many iterations to get reader feedback, pivot until you have
the right book and build traction once you do.

This work is licensed under a Creative Commons

Attribution-NonCommercial-ShareAlike 4.0 International License
Tweet This Book!
Please help Peter Prevos by spreading the word about this book on
Twitter!
The suggested tweet for this book is:
How do you create value from data to support business strategy?
This book by @pprevos provides a framework for strategic data
science that goes beyond the hype #datascience
https://leanpub.com/strategic_data_science/
Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Preface to second edition . . . . . . . . . . . . . . . . . . iv

Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . v

1. What is Data Science? . . . . . . . . . . . . . . . . . . . 1

1.1 Data-Driven Organisation . . . . . . . . . . . . . . 7
1.2 The Data Revolution . . . . . . . . . . . . . . . . . 10
1.3 The Elements of Data Science . . . . . . . . . . . 11
1.4 The Purpose of Data Science . . . . . . . . . . . . 16

2. Good Data Science . . . . . . . . . . . . . . . . . . . . . 18

2.1 Data Science Trivium . . . . . . . . . . . . . . . . 19
2.2 Useful Data Science . . . . . . . . . . . . . . . . . 20
2.3 Sound Data Science . . . . . . . . . . . . . . . . . 28
2.4 Aesthetic Data Science . . . . . . . . . . . . . . . . 36
2.5 Best-Practice Data Science . . . . . . . . . . . . . 45

3. Strategic Data Science . . . . . . . . . . . . . . . . . . . 47

3.1 The Data Science Continuum . . . . . . . . . . . . 47
3.2 Collect data . . . . . . . . . . . . . . . . . . . . . . 49
3.3 Descriptive Statistics . . . . . . . . . . . . . . . . . 52
3.4 Diagnostics . . . . . . . . . . . . . . . . . . . . . . 55
3.5 Predicting the Future . . . . . . . . . . . . . . . . . 57
3.6 Prescribe Action . . . . . . . . . . . . . . . . . . . 63
3.7 Towards a Data-Driven Organisation . . . . . . . 64
CONTENTS

4. The Data-Driven Organisation . . . . . . . . . . . . . . 66

4.1 People . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Systems . . . . . . . . . . . . . . . . . . . . . . . . 76
4.3 Process . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.4 The Limitations of Data Science . . . . . . . . . . 86

5. References . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Preface
I am not a data scientist. These might be strange words to open a
book about data science, so please allow me to explain. This book is
the result of a twenty-five-year career in civil engineering, building
and managing structures in Europe, Africa, Asia and Australia.
Most of my tasks involve managing and analysing large amounts
of data. Cost estimates, volume calculations, modelling river flows,
structural calculations, Monte Carlo simulations and many other
types of number crunching are integral to my work as an engineer.

My journey towards what we now call data science started at

university. When studying engineering in the Netherlands, I wrote
computer code in the Pascal and BASIC languages. I loved spending
time in the computer lab and write software to solve technical
problems. The dean advised me to switch from civil engineering
to computer science, but I enjoyed writing software to solve engi-
neering problems, not for the sake of it, so I did not heed the advice.

In my first job as a civil engineer, my company introduced the

now-defunct Lotus 123 spreadsheet. When first using this package,
I thought it was the best thing since sliced bread. Graphical output
and managing data were complex tasks when writing code in those
days. The allure of the spreadsheet was the ability to combine
inputs with computer code and show the results in text or graphs,
all in one convenient file.

Over the next two decades, I have written hundreds of spreadsheets

to solve a myriad of engineering problems. I even developed a
‘jungle’ of interconnected spreadsheets to manage the logistics of
a large river engineering project in Bangladesh. The complexity of
this task took me to beyond limits of what spreadsheets can achieve.
Preface ii

Throughout my career, I had many nightmarish experiences trying

to reverse engineer spreadsheets to figure out how they work, even
ones I wrote myself. The combination of data, code and output that
I loved at the start of my career was now a source of frustration.
My love affair with the venerable spreadsheet ended when writing
my doctoral dissertation. Excel was incapable of helping me with
complex statistics such as structural equation modelling or network
analysis. A colleague suggested looking into this new thing called
‘Data Science’. I decided to learn how to write code in R, a spe-
cialised computer language for statistical analysis. The R language
is like a Swiss army chainsaw for engineers, with capabilities that
far exceed anything a spreadsheet can do.
I now manage the data science function for a water utility in
regional Australia. Through my experience with practical data
analysis and expertise in management, I have developed a strategic
approach to data science. I have published and presented my views
on strategic data science at conferences in Australia and New
Zealand. This book is an expanded version of an article I wrote for
the journal of the Australian Water Association, which became one
of the most downloaded papers. Lifting the Big Data Veil describes
a back-to-basics approach on how to maximise the value we can
extract from data assets.¹ This book dives deeper into the principles
of data science first presented in this paper.
I write this book from the perspective of an engineer and a social
scientist, but the same principles are valid for any field of human
endeavour. All professionals and scientists rely on data to make
decisions. Data science provides a systematic approach to making
better business decisions and discovering new patterns in society
or nature.
My approach in this book goes back to the basics of what it means
to create value from data. This book is not a treatise about machine
¹Prevos, P. (2017). Lifting the ‘Big Data’ Veil. Creating Value through Applied
Data Science. Water E-Journal, 2(1), 1–5. DOI 10.21139/wej.2017.008.
Preface iii

learning, mathematics or developing software, but a practical guide

to strategically and systematically using data to create a better
world. This book is pragmatic because it doesn’t dwell on the future
promises of machine learning, artificial intelligence or quantum
computing. The framework in this book is inspired by my current
and desired practice as an engineer and social scientist, with a data
science responsibility and best-practice in management.

The information in this book is by no means the final say on the

topic of strategic data science. This book is opinionated in that the
approach it describes is only one way of looking at the subject.
This book is not an academic dissertation that evaluates various
methods to the science and craft of analysing data. My experience
and knowledge profoundly influence my views on the topic as a
civil engineer and training in the humanities. My objective for this
book is to encourage the reader to think beyond the algorithms and
shiny data tools and seek a deeper understanding of what it is you
want to achieve.

My primary objective is to introduce the reader to a wide range

of considerations to improve the way they manage data and to
provoke curiosity to research the topic in more depth. My objective
is to inspire data professionals and managers to worry less about
the technical details and think strategically about extracting value
from data.

The raw text of this book and R code for the data visualisations is
available on GitHub. I encourage anyone who discovers mistakes
or likes to enhance the information in these pages to contact me.
The LeanPub publishing system provides flexible opportunities
to publish new versions when updated information is available.
Anyone purchasing this book through LeanPub can register to be
informed of future editions.
Preface iv

Preface to second edition

The second edition of this book contains many grammar fixes.

Thanks to David Smith and Catherine Cousins for spotting these
mistakes.

Interestingly, the machine learning program I used for checking the

text did not identify many of these mistakes in the first version. This
experience strengthens one of the points in this book, which is that
artificial intelligence is not a replacement for natural intelligence.
Acknowledgements
My career has been dynamic and varied but analysing data has
been the one constant in my career. My toolkit to convert data
to actionable intelligence has evolved significantly over the past
twenty-five years. My former and current colleagues helped me to
accumulate new approaches and skills, which led me to write this
book.
In my first job as an engineer, Wim van Vliet in Johannesburg was
a significant influence on my development as an engineer. Through
him, I learned how to convert my theoretical knowledge into a
systematic approach to solving real problems.
René Zekveld was my manager at Boskalis when I worked on
marine engineering projects in Hong Kong and Bangladesh. René
taught me tricks of the trade on how to interpret data to achieve
sound business outcomes. He emphasised the importance of the
relationship between reality and data. Our long discussions about
data showed me that actionable intelligence and insight are essen-
tial outcomes.
After my international career, I moved back to the Netherlands
and spent time with Rijkswaterstaat, the government agency re-
sponsible for keeping the country from flooding. Working for
this organisation exposed me to probabilistic approaches in cost
estimation and project management. We regularly worked with
mathematicians, whose names are lost in time, to help us grasp the
complexity of our analysis. Working for this organisation advanced
my skills in using mathematics to create value from data.
At the start of the new millennium, I started my current job at
Coliban Water, a water utility in regional Australia. Brad Dole,
Acknowledgements vi

a former IT manager, has been influential in how I now work

with data. As an IT manager, Brad provided me with the freedom
to maximise the capabilities of the company hardware. Brad also
has a razor-sharp insight into data and how to convert business
problems into code. Under Brad’s guidance, we developed several
in-house software solutions, which took my software development
skills beyond a mere hobby.

Jenny Fogarty is my current colleague and the data architect at Col-

iban Water. We have worked on developing software and reporting
mechanisms for many years. Through her expertise, I learned
everything I know about managing data. It was on her advice five
years ago that I started researching the topic of data science. Not
only has her advice helped me to complete my dissertation about
customer centricity, but it has also transformed the way I view my
profession.

No vision of data science can ever be implemented without people

doing the work. Gary Schurr has over the past few years been great
at finding innovative ways to report information. His critical helps
to straighten out some of my more impulsive ideas.

I started my formal journey into data science by doing a range of

courses on the Coursera website. The data science specialisation by
John Hopkins University taught me the principles of how to use
the R language for statistical computing. The inspiring lectures by
Roger Peng, Jeff Leek and Brian Caffo helped me to think more
systematically in how I deal with data.

My manager David Sheehan often encourages me to develop new

ideas and ‘conquer the world’. This book is just another small step
in that direction. It would not have existed without the freedom
David provides me to shape the data science function at Coliban
Water and the broader water industry in Australia.

Lastly, I thank all my colleagues at Coliban Water and the broader

Australian water industry who indulge me in my geekiness and
Acknowledgements vii

enthusiasm to find data solutions to improve the way we service

our customers. This book is the direct result of the positive feedback
from the people that have attended my conference presentations
and read my publications and articles on my Lucid Manager
(lucidmanager.org) website.
1. What is Data Science?
The activity of analysing data is as old as human culture. The
earliest known form of writing is not an epic poem or religious
text, but data. The Ishango bone is an engraved fibula of a baboon
which was carved in central Africa 20,000 years ago. Some scholars
hypothesised that the carvings represent an early number system as
it lists several prime numbers, while others believe it to be a calen-
dar. Some researchers dismiss these ideas and believe the markings
merely improve grip when using the bone as a club. Whatever their
purpose, the groupings of the markings are distinctly mathematical
(Figure 1).¹

Figure 1: Markings on the Ishango Bone

Ancient cultures around the world collected data by observing na-

ture and the stars to predict when they needed to move camp, start
sowing crops, hunt seasonal animals or to obtain whatever other
knowledge required for survival. These proto-scientific methods
¹Pletser, V. (2012). Does the Ishango Bone Indicate Knowledge of the Base 12? An
Interpretation of a Prehistoric Discovery, the First Mathematical Tool of Humankind.
Eprint ArXiv:1204.1019.
What is Data Science? 2

were the first attempts at science as these early researchers collected

data to explain the world in logical terms. These primitive forms of
science helped these people to understand their world and control
their destiny, which is precisely what contemporary science seeks
to achieve.

Mathematics was an integral part of ancient civilisations. Sumeria,

Egypt, Rome and other advanced ancient civilisations used math-
ematics to manage their society and build their elaborate cities.
The origins of civilisation as we now know it lies in Mesopotamia,
current-day Iraq. Archaeologists have excavated thousands of clay
tablets that record their day to day activities such as land sales,
delivery of goods and other commercial transactions. Around that
same time in Pharaonic Egypt, the first census took place, recording
demographic data about its inhabitants.² These examples show that
collecting data and using it to control and improve our world is an
ancient human activity.

This time was also a period of the first significant mathematical

discoveries and inventions. Mathematics was, however, more than
a language to model the world. To the great Greek mathematician
Pythagoras, numbers possessed meaning beyond their ability to
describe quantity. In these early days of intellectual exploration,
divination was the most popular method to predict the future.
Astrologers mapped the skies or studied the entrails of a bird to
find a relationship between these patterns and their world. In these
divination systems, mathematics was practised as a tool to manage
society through engineering and bookkeeping, not as a tool to
describe the world.

The scientific revolution of the seventeenth century replaced div-

ination with a mathematical approach to understanding the world.
Since the work of René Descartes, mathematics is a method to
²Kelleher, J.D., & Tierney, B. (2018). Data science. Cambridge, Massachusetts:
The MIT Press.
What is Data Science? 3

describe the world and to predict its future.³ This revolution in

how we perceive the world mathematically enabled the industrial
revolution. Early technology enhanced our physical capabilities
with machines, while modern technology improves our minds with
computers. Machines make us stronger and faster, and their devel-
opment revolutionised society during the first industrial revolution.
Computers enhance aspects of our mental abilities, and we are in
the middle of a second industrial revolution which is not fuelled
with oil and coal but with data.

The idea that data can be used to understand the world is thus
almost as old as humanity itself and has gradually evolved into
what we now call data science. We can use some basic data science
to review the development of this term over time.

Figure 2: Frequency of the bi-gram ‘data science’ in literature and Google searches
in the percentage of highest occurrence.

³Davis, P.J., & Hersh, R. (1990). Descartes’ Dream. The World According to
Mathematics. London: Penguin.
What is Data Science? 4

The combination of the words data and science might seem rel-
atively new, but the Google N-gram Viewer shows that this ‘bi-
gram’ has been in use since the middle of the last century. An n-
gram is a sequence of words, with a bi-gram being any combination
of two words. Google’s n-gram viewer is a searchable database of
millions of scanned books published before 2008. This database is
a source for predictive text algorithms as it contains a fantastic
amount of knowledge about how people use various languages.⁴

The n-gram database shows that the term data science emerged in
the middle of the last century when electronic computation became
a topic of study. In those days, the discipline was a science about
storing and manipulating data. The current definition has drifted
away from this initial academic activity to a business activity.

Judging by another Google database, data science started its jour-

ney from obscurity to becoming the latest business fad only a few
years ago. The Google Trends database shows the frequency of
search terms over time. Google Trends reveals a steady increase
in the popularity of data science as a search term, starting in 2013.⁵

Figure 2 visualises these two trends. The horizontal axis shows the
years from 1960 until recently. The vertical axis shows the relative
number of occurrences compared to the maximum, which is how
Google report search numbers. In an absolute sense, the number of
occurrences in books was much lower than current search volumes.
While the increase in attention has steeply risen since 2012, the
term started its journey towards business buzzword in the 1960s.
Although we can speak of a recent hype, in essence, data science is
a slow evolution with a recent spike in interest.

The expectations of the benefits of data science are very high.

Business authors position data science, and its natural partner ‘big
data’, as a panacea for all societal problems and a means to increase
⁴Google Books Ngram Viewer. Retrieved 25 January 2019.
⁵Google Trends. Retrieved 25 January 2019.
What is Data Science? 5

business profits.⁶ In a 2012 article in Harvard Business Review,

Davenport and Patil even proclaimed data scientist the “sexiest job
of the 21st century”.⁷ Who would not want to be part of a new
profession with such enticing career prospects? It is not a stretch to
hypothesise that their article was one of the causes of the increased
search volume reported by Google.
The recent popularity of data science as a business activity suggests
that it is just a fancy way of saying business analysis. In my talks
about data science in Australia and New Zealand, I regularly meet
fellow managers who are sceptical about the proclaimed wizard-
like abilities of data scientists and the unbounded promises of
machine learning. Much of the data science promise relates to
the success stories of Internet corporations such as Google and
Facebook and many other smaller players in the digital economy.
For these organisations, data science is a core competency as their
value proposition centres around data.
For organisations that deliver physical products or non-digital
services, data science is about improving how they collect, store and
analyse data to extract more value from this resource. The objective
of data science is not the data itself but is closely intertwined with
the strategic goals of the organisation. These objectives broadly
range from increasing the return to shareholders to providing
benefit to society overall. Whatever the kind of organisation you
are in, the purpose of data science is to assist managers to change
reality into a more desirable state. A data scientist achieves this
by measuring the current and past states of reality and using
mathematical tools to predict a future state.
The term data science is unfortunate in the way it is now used
because it is paradoxically not a science about data. A data scientist
is not somebody who researches the properties of data. Other
⁶Clegg, B. (2017). Big Data: How the Information Revolution Is Transforming
Our Lives. Icon Books.
⁷Davenport, T.H., & Patil, D.J. (2012). Data scientist: The sexiest job of the 21st
century. Harvard Business Review, 90(10), 70–76.
What is Data Science? 6

definitions see data scientists as mathematicians and computer

scientists that invent new ways of analysing data. More commonly,
data science is closely related to business outcomes.

Data science is a systematic and strategic approach to using data

to solve practical problems. The problems of the data scientist
are practical because pure science has a different objective to
business. A data scientist in an organisation is less interested in
a generalised solution to a problem and focuses on improving how
the organisation achieves its goals. Perhaps the combination of the
words data and science should be reserved for academics.

There are some signals that the excitement of the past few years is
waning. Data science blogger Matt Tucker has declared the death
of the data scientist.⁸ For many business problems, the hardcore
methods of machine learning result in over-analysing the issue.
Tucker provides an anecdote of a group of data scientists who
spend a lot of time fine-tuning a complex neural network. The
data scientists gave up when a young graduate with expertise in
the subject-matter used a linear regression that was more accurate
than their neural network.

The negative chatter on the internet about data science as a business

discipline might imply that the hype is receding. We should, how-
ever, not throw the baby out with the bathwater. The recent interest
in analysing data has raised the stakes in how organisations use this
valuable resource. Even after the inflated expectations recede, data
science as a profession has a useful contribution to make. All data,
big and small, is a resource to improve how organisations perform.

This book looks at data science as the strategic and systematic ap-
proach to the fine art of analysing data to solve business problems.
This conceptualisation of data science is not a complete definition.
Computational analysis of data is also practised as a science in itself
⁸Tucker, M. (2018). The Death of the Data Scientist. Retrieved 9 February 2019
from Data Science Central.
What is Data Science? 7

and as a scientific method for research in many areas. This book is

written from a business perspective, and these other uses on data
science are not further considered.

1.1 Data-Driven Organisation

The previous section showed that data science is not necessarily

a hype, but a strategic and systematic approach to using data.
Using data in organisations is also called business analytics or
evidence-based management. There are also specific approaches,
such as Six-Sigma, that use statistical analysis to improve business
processes. Many advocates of data science claim that the old and
new approaches are different. Most definitions of data science
focus on pattern recognition using large sets of data through
machine learning.⁹ How does data science relate to its predecessor
buzzwords? To understand this difference we need to explore the
early history of using data in business.

The idea that management can be science is just over a cen-

tury old. Frederick Taylor was an American engineer who was
dissatisfied with how factories were managed. He was a hands-
on engineer who spent much time on the factory floor. Taylor
noticed how workers used rules-of-thumb instead of analysing
problems systematically. He writes in The Principles of Scientific
Management (1911) how he improved the process of manually
loading massive lumps of iron at the Midvale Steel Company by
measuring processes and analysing the data.¹⁰

Although Taylor revolutionised the way we manage organisations,

he dispised labourers. Taylor believed that it “would be possible
to train an intelligent gorilla to become more efficient” than a
⁹Kelleher & Tierney (2018).
¹⁰Taylor, F.W. (1997). The Principles of Scientific Management. Mineola, N.Y:
Dover Publications.
What is Data Science? 8

factory worker. His quest for scientific management was driven

by an urge to remove power from the workforce and look at
business processes in an abstract mathematical sense. His work
was controversial in his own time as it was the subject of a formal
government enquiry. This background about Taylor is not just a bit
of trivia but a valuable lesson about ensuring to include a human
dimension in what we analyse. The positive legacy of Taylor is
that he planted the seed for a scientific approach to managing an
organisation. All methods share his ideal to use data to prevent
biases in management.

Managers are faced with deciding what to do next in uncertain

environments and often use their experience and intuition to de-
termine the next course of action instead of data and logic. While
experience and intuition are highly valuable, our minds are prone
to biases and non-rational thinking. Our rationality is not unlimited
but bounded by factors outside of our control. The amount of
information, time to solve a problem and our mental capacity are
limited. Our brains are wired to quickly recognise patterns in nature
because it helps us in our daily lives. Mental shortcuts, the rules-
of-thumb despised by Taylor, help us to make fast decisions in
emergencies, but they can also lead to sub-optimal outcomes.

The world of business is not something we have necessarily evolved

to navigate, and we are thus not very good at interpreting large
amounts of abstract data. Because our minds are programmed to
recognise patterns, we often see regularities where there are none,
which psychologists call pareidolia. This condition causes us to
recognise animals in clouds, see the image of Jesus in a piece of
toast or a face on Mars. Pareidolia serves us well because it en-
ables graphical communication, but it becomes a hindrance when
analysing large sets of data. Interestingly, neural networks can
also be trained to experience pareidolia. When manipulating the
settings of image-recognition software, computers can be taught to
recognise images in random data and effectively hallucinate a new
What is Data Science? 9

reality.¹¹

Besides inherent biases through the limits of our rationality, so-

cial circumstances can also prevent us from optimising decisions.
Group-think and office politics are often strong drivers of decisions
in organisations. Social belonging is a strong motivator for our
behaviour and is one of the major driving forces behind advertising.
Asch’s conformity experiment illustrates how strong these social
biases can be. Solomon Asch demonstrated that even when people
are fully aware of the rational answer to a simple question, they
will in most instances yield their opinion to match that of the group,
even when it is clearly the wrong choice.¹²

One of the greatest revolutions in human thinking is the 15th-

century Copernican twist. From our limited perspective, the earth
seems flat and the sun and moon revolve around us. When Coper-
nicus looked through a telescope to amplify his naked-eye obser-
vations, a new reality emerged with a better model of our solar
system. What we learned from Copernicus is that we need to
enhance our perception and thinking skills with technology to draw
correct conclusions. Data science is to business what the telescope is
to astronomy. Sound analysis of data helps us to remove our natural
biases and replace our rules-of-thumb with logic.

Just like a lever makes us physically stronger to lift the world,

the tools of data science make us mentally stronger to understand
and control the world. While the uncertainties of the realities
of business can never be eliminated, evidence-based management
ensures that managers make decisions based on the best available
data. Data science is the toolkit that assists managers to base their
decisions on evidence. Using the principles of data science will
improve the way managers decide between alternative courses of
action.
¹¹Mordvintsev, A., Olah, C. & Tyka, M.. (2015). Inceptionism: Going Deeper into
Neural Networks. Retrieved 15 February 2019.
¹²Asch Conformity Experiment (Youtube). Downloaded 14 February 2019.
What is Data Science? 10

Using a scientific approach to data is, however, not a simple road to

success. Data science is a human activity that encompasses all the
above biases and limitations. The results of data science are also
not ethically neutral and require a moral perspective to ensure that
no harm is done. The key to minimising these biases is to use a
systematic approach.

1.2 The Data Revolution

Since Taylor’s first writings, businesses and non-profit organisa-

tions have sought to become driven by evidence to reduce uncon-
scious bias in their decisions. Although data science is merely a
new term for something that has existed for decades, some recent
developments have created a watershed between the old and new
ways of doing business. The difference between traditional business
analysis and the new world of data science is threefold.

Firstly, businesses have much more data available than ever be-
fore. The move to electronic transactions means that almost every
process leaves a digital footprint. Collecting and storing this data
has become exponentially cheaper than in the days of pencil and
paper. Many organisations collect this data without maximising the
value they extract from it. After the data is used for its intended
purpose, it becomes ‘dark data’, stored on servers but languishing
in obscurity. This data provides opportunities to optimise how an
organisation operates by recycling and analysing it to learn about
the past to create a better future.

Secondly, the computing power that is now available in a tablet

was not long ago the domain of supercomputers. Piotr Luszczek
showed that an iPad-2 matches the performance of the world’s
fastest computer in 1985.¹³ The affordability of vast computing
¹³Larabel, M. (2012). Apple iPad 2 As Fast As The Cray-2 Super Computer.
Retrieved 4 February 2019 from Phoronix.
What is Data Science? 11

power enables even small organisations to reap the benefits of

advanced analytics.

Lastly, complex machine learning algorithms are freely available

as Open Source software, and a laptop is all that is needed to
implement sophisticated mathematical analyses. The R language
for statistical computing and Python are potent tools that can
undertake a vast array of data science tasks such as complex
visualisations and machine learning. These languages are ‘Swiss
army chainsaws’ that can tackle any business analysis problem.
Part of their power lies in the healthy communities that support
each other in their journey to mastering these languages.

These three changes have caused a revolution in how we create

value from data. The barriers to entry for even small organisations
to leverage information technology are very low. The only hurdle
to take is to make sense out of the fast-moving developments and
follow a strategic approach instead of chasing the hype.

This revolution is not necessarily only about powerful machine

learning algorithms, but about a more scientific way of solving
business problems. The definition of data science in this book is not
restricted to machine learning, big data and artificial intelligence.
These developments are essential aspects of data science, but they
do not define the field.

1.3 The Elements of Data Science

Now that we have defined data science within the context of

managing a business, we can start describing the elements of data
science. The best way to unpack the art and craft of data science is
Drew Conway’s often-cited Venn diagram cite (Figure 3).¹⁴
¹⁴Conway, D. (2010). The data science Venn diagram.. Downloaded 27 January
2019.
What is Data Science? 12

Conway defines three competencies that a data scientist, or a data

science team as a collective, need to possess. The diagram positions
data science as an interdisciplinary activity with three dimensions:
domain knowledge, mathematics and computer science. A data
scientist is somebody who understands the subject matter under
consideration in mathematical terms and writes computer code to
solve problems.

Figure 3: Conway’s data science Venn Diagram.

1.3.1 Domain Knowledge

The most significant skill within a data science function is domain

knowledge. While the results of advanced applied mathematics
such as machine learning are impressive, without understanding
the reality that these models describe, they are devoid of meaning
and can cause more harm than good. Anyone analysing a problem
needs to understand the context of the issues and the potential
solutions. The subject of data science is not the data itself but the
reality this data describes. Data science is about things and people
in the real world, not about numbers and algorithms.
A domain expert understands the impact of any confounding
variables on the outcomes. An experienced subject-matter expert
What is Data Science? 13

can quickly perform a sanity check of the process and results of

the analysis. Domain knowledge is essential because each area of
expertise uses a different paradigm to understand the world.

Each domain of human enquiry or activity has different method-

ologies to collect and analyse data. Analysing objective engineering
data follows a different approach to subjective data about people
or unstructured data in a corpus of text. The analyst needs to be
familiar with the tools of the trade within the problem domain.
The earlier-mentioned example of a graduate professional beating
a team of machine learning experts with a linear regression shows
the importance of domain knowledge.

Domain expertise can also become a source of bias and prevent

innovative ways of looking at information. Solutions developed
through systematic research can contradict long-held beliefs about
a specific topic that are sometimes hard to shift. Implementing data
science is thus as much a cultural process as it is a scientific one,
which is the topic of chapter four.

1.3.2 Mathematical Knowledge

The analyst uses mathematical skills to convert data into actionable

insights. Mathematics consists of pure mathematics as a science
in itself and applied mathematics that helps us to solve problems.
The scope of applied mathematics is broad, and data science is
opportunistic in choosing the most suitable method. Various types
of regression models, graph theory, k-means clustering, decision
trees and so on, are some of the favourite tools of a data scientist.
The creative application of complex applied mathematics is one of
the two distinguishing factors between traditional business analysis
and data science.

Combining subject-matter expertise with mathematical skills is

the domain of traditional research and analysis. The notion of
What is Data Science? 14

conventional research is, however, evolving towards using the

principles of data science by using reproducible computer code
and sharing the source data through websites such as FigShare
(figshare.com).

Numbers are the foundations of mathematics, and the craft of

quantitative science is to describe our analogue reality into a model
that we can manipulate to predict the future. Not all mathematical
skills are necessarily about numbers but can also revolve around
logical relationships between words and concepts. Contemporary
numerical methods help us to understand relationships between
people, the logical structure of a text and many other aspects
beyond the realm of traditional numeric analysis.

1.3.3 Computer Science

Not that long ago, most of the information collected by an organ-

isation was stored on paper and archived in copious volumes of
arch lever files. Analysing this information was an arduous task
that involved many hours of transcribing information to a format
that is useful for analysis.

In the twenty-first century, almost all data is an electronic resource.

To create value from this resource, data engineers extract it from a
database, combine it with other sources and clean the data before
analysts can make sense of it. This requirement implies that a data
scientist needs to have computing skills. Conway uses the term
hacking skills, which many people interpret as negative. Conway
is, however, not referring to a hacker in the sense of somebody
who nefariously uses computers, but in the original meaning of
the word of a developer with creative computing skills. The core
competency of a hacker, developer, coder or whatever other term
might be preferable, is algorithmic thinking and understanding the
logic of data structures. These competencies are vital in extracting
What is Data Science? 15

and cleaning data to prepare it for the next step of the data science
process.

The importance of hacking skills for a data scientist implies that we

should move away from point-and-click systems and spreadsheets
and instead write code in a suitable programming language. The
flexibility and power of a programming language far exceed the
capabilities of graphical user interfaces and leads to reproducible
analysis, discussed in chapter two.

The mathematical interpretation of reality needs to be translated

to computer code. One of the factors that spearheaded data science
into popularity is that the available toolkit has grown substantially
in the past ten years. Open source computing languages such as R
and Python are capable of implementing complex algorithms that
were previously the domain of specialised software and supercom-
puters. Open Source software has accelerated innovation in how
we analyse data and has placed complex machine learning within
reach of anyone who is willing to make an effort to learn the skills.

Conway defines the danger zone as the area where domain knowl-
edge and computing skills combine, without a good grounding in
mathematics. Somebody might have sufficient computing skills to
be pushing buttons on a business intelligence platform or spread-
sheet. The user-friendliness of some analysis platforms can be
detrimental to the outcomes of the analysis because they create
the illusion of accuracy. Point-and-click analysis hides the inner
workings from the user, creating a black-box result. Although the
data might be perfectly structured, valid and reliable, a wrongly-
applied analytical method leads to useless outcomes.

1.3.4 The Unicorn Data Scientist?

Conway’s diagram is often cited in the literature on data science.

His simple model helped to define the craft of data science. Other
What is Data Science? 16

data scientists have proposed more complex models but they all
originate with Conway’s basic idea. A quick visit on the Duck-
duckgo search engine will reveal several variants.

The diagram illustrates that the difference between traditional

research skills or business analytics lies in the ability to understand
and write code. A data scientist understands the problem they
seek to resolve, they have the mathematical expertise to analyse
the problem, and they possess the computing skills to convert this
knowledge into outcomes.

It could be argued that the so-called skills are missing from this
picture. However, communication, managing people, facilitating
change and so on, are competencies that belong to every profession
who works in a complex environment, not just the data scientist.

Some critics of this idea point out that these people are unicorns.
Data scientists that possess all these skills are mythical employees
that don’t exist in the real world. Most data scientists start from
either mathematics or computer science, after which it is hard to
become a domain expert. This book is written from the point of
view that we can breed unicorns by teaching domain experts how to
write code and, where required, enhance their mathematical skills.

1.4 The Purpose of Data Science

In summary, the promises of data science within organisations have

gained a lot of popularity over the past six years. The downside of
this popularity is that self-proclaimed futurists have exaggerated
the benefits of a strategic and systematic approach to analysing
data. To obtain value from this new approach to using data requires
a pragmatic approach beyond the hype. For the vast majority
of organisations, data science will look very differently from the
digital utopia portrayed in popular publications.
What is Data Science? 17

This chapter defines data science as the strategic and systematic

use of data to create value for organisations or society overall. The
purpose of using data to improve how organisations perform is
to reduce bias in decisions. The original objections that Frederick
Taylor held against rules-of-thumb more than a century ago still
stands. Computational analysis of data is a valuable tool to achieve
this reduced bias in deciding about future courses of action.

Data science is an interdisciplinary activity that combines do-

main knowledge with competencies in mathematics and computer
science. The data revolution of the past decades has caused an
exponential increase in available data, computing capabilities and
open source software. Data science is paradoxically not a science
about data but a scientific way to use data to influence reality pos-
itively. Expertise about the reality under consideration, or domain
knowledge, drives data science. Mathematics and computer science
are the tools that enable a deeper understanding of our reality and
help us to optimise our decisions.

Now that we have an idea of what data science is and what it

consists of, we need to define what good data science looks like. The
following chapter expands on this description of data science by
presenting a normative model of data science. This model defines
best practice as the useful, sound and aesthetic analysis of data.
2. Good Data Science
Data is sometimes called the “new oil”, a source of new-found
wealth that is mined from the depths of corporate and government
archives. Some accountants are so excited about the potential value
of data that they account for it in the same way as a physical
asset. While it is indeed true that data can increase the value of
an organisation, this resource has no intrinsic value. Just like oil,
data needs to be mined and has to be refined to the right quality.
Data needs to be transported through information networks before
it can be used to create new value. The value of data is not in
the information itself but occurs during the transformations it
undergoes.

The analogy between data and oil is only partially correct in that
data is a renewable resource. The same data can be used many
times for a sometimes originally unintended purpose. The ability
to use data for more than one goal is one of the reasons data
science has gained popularity around board tables. Senior managers
are seeking ways to extract value from so-called ‘dark data’. Data
scientists use these forgotten data sources to create new knowledge,
make better decisions and spawn innovations.

The question that arises from this introduction is how to manage

and analyse data so it can become a valuable resource. This chapter
presents a normative model to create value from data using three
basic principles derived from architecture. This model is useful for
data scientists as an internal check to ensure that their activities
maximise value. Managers can use this model to assess the out-
comes of a data science project without having to understand the
mathematical intricacies of the craft and science of analysis.
Good Data Science 19

2.1 Data Science Trivium

Although data science is a quintessential twenty-first-century ac-

tivity, to define good data science, we can find inspiration in a
Roman architect and engineer who lived two thousand years ago.
Vitruvius is immortalised through his books About Architecture,
which inspired Leonardo Da Vinci to draw his famous Vitruvian
man. Vitruvius wrote that an ideal building must exhibit three
qualities: utilitas, firmitas and venustas, or usefulness, soundness
and aesthetics.

Figure 4: The principles of good data science.

Buildings must have utility so they can be used for their intended
purpose. A house needs to be functional and comfortable; a theatre
needs to be designed so that everybody can see the stage. Each type
of building has its functional requirements. Secondly, buildings
must be sound in that they are firm enough to withstand the forces
that act upon them. Last but not least, buildings need to be aesthetic.
In the words of Vitruvius, buildings need to look like Venus, the
Roman goddess of beauty and seduction.
The Vitruvian rules for architecture can also be applied to the
products of data science.¹ Great data science needs to have utility;
¹Lankow, J., Ritchie, J., & Crooks, R. (2012). Infographics: The Power of Visual
Storytelling. Hoboken, N.J: John Wiley & Sons, Inc.
Good Data Science 20

it needs to be useful to create value. The analysis should be sound

so it can be trusted. The products of data science also need to be
aesthetic to maximise the value they provide to an organisation
(Figure 4).

2.2 Useful Data Science

How do we know that something is useful? The simple, but not

very illuminating answer is that when something is useful, it has
utility. Some philosophers interpret utility as the ability to provide
the greatest good for the highest number of people. This definition
is quite compelling, but it requires contextualisation. What is right
in one situation might not be so beneficial in another.
The highest number of people is open to interpretation. Is some-
thing only useful when it benefits all of humanity or can it also be
useful when it helps just one person? The requirement to include
the highest number of people in our definition of usefulness might
work well for government organisations. It is not so clear in corpo-
rations that seek to maximise the benefits to their shareholders.
Whether something is useful or useless depends on the context in
which it is applied, but also on the set of values that it is judged by.
Defining usefulness in generic terms will prove to be an impossible
quest because of the dependence on the context and the relevant
value system. For example, to Greenpeace, analysing data from
fracking activities will have different usefulness as it does to a gas
exploration company. The same data can satisfy different types of
merit depending on context.
These philosophical deliberations aside, defining usefulness for
organisations is more straightforward because we apply a prag-
matic approach. Usefulness for organisations is the extent to which
something contributes to their strategic or operational objectives. If
the result of a data science project is unable to meet this criterion,
Good Data Science 21

then it is strictly speaking useless. As a civil engineer and social

scientist, I could spend many hours analysing the vast amounts
of data collected by my organisation. Dredging the data to find
something of value might be an exciting way to waste time, there
is also a significant risk of finding fool’s gold instead of valuable
nuggets of wisdom. The first step that anyone working with data
should undertake before starting a project is to define the business
problem that needs solving.

This book follows a pragmatic and perspectivist view of usefulness.

For a data science strategy to be successful, it has to facilitate the
objectives of the organisation. Data scientists are opportunistic in
the approach they use to resolve problems. Perspectivism implies
that the same data can be used for different issues, depending on the
perspective you take on the available information and the problem
at hand.

After digesting a research report or viewing a visualisation, man-

agers ask themselves: “What do I do differently today?” Usefulness
in data science depends on the ability of the results to empower
professionals to influence reality positively. In other words, the
result of data science should either comfort management that
objectives have been met, or they provide actionable insights to
resolve existing problems or prevent future ones.

Providing actionable intelligence is only a narrow scope of works

for data scientists. The concept of usefulness in business needs to
be extended beyond this short term and one-dimensional view.
Business scholar Bernard Jaworski classified the results of research
into two types to help us make sense of how theory relates to
practice.² Some knowledge is suitable for action, which is the much
sought-after actionable intelligence. Other research doesn’t lead to
action but inspires deeper thinking about managerial practice. Data
science can direct action, but it can also motivate innovation by
²Jaworski, B.J. (2011). On managerial relevance. Journal of Marketing, 75(4),
211–224. doi 10.1509/jmkg.75.4.211.
Good Data Science 22

providing a more profound understanding of the current reality.

The results of data science should either stimulate action or inspire
contemplation. While the relevance of taking action is self-evident,
reflection is not often recognised as a beneficial managerial impact.

Figure 5: Reality, Data, Information, Knowledge Pyramid.

For data science to provide actionable intelligence, the raw data

needs to be converted to knowledge following a standardised
workflow. The well-known DIKW Pyramid (Data, Information,
Knowledge and Wisdom) explains how data produces a useful
analysis. The source of the original version of this model is lost
in time as a multitude of authors has used it without citation. The
basic principle of the hierarchy is that to obtain wisdom you need to
have relevant knowledge, which derives from information, which
in turn consists of the conclusions drawn from the data. Various
versions of the model have been proposed, with slightly different
terminology and interpretations.

The version in this book is modified to understand better how

to create useful data science (Figure 5). Firstly, wisdom no longer
forms part of the model because this concept is too nebulous to be
useful. Anyone seeking wisdom should study philosophy or prac-
tice religion as data science is unable to provide this need. Secondly,
the bottom of the pyramid needs to be grounded in reality. The
standard DIKW model ignores the reality from which the data is
collected that creates the information and knowledge used to make
business decisions. The second addition to the traditional model is
Good Data Science 23

a feedback loop from knowledge to the real world. The purpose of

data science is to enhance the knowledge that professionals use to
influence reality by converting data into information.

2.2.1 Reality

Useful data science positively influences reality by collecting data,

creating information and increasing our knowledge about and
understanding of reality. This knowledge is useful when it changes
the way we perceive reality to innovate the way we do things, and
when it enables better operational or strategic decisions. If data
science becomes abstracted from the world it seeks to understand
or influence, it loses its power to be valuable.

This reality of data science can be either physical or social, each

of which requires a different paradigm to describe the world. Our
physical reality can be measured with almost arbitrary precision.
We can measure size, weight, chemical composition, time and so
on, with high validity and reliability.

The social world can also be summarised in numbers, but these

measurements are almost always indirect. We cannot read people’s
minds. When we want to know how somebody feels about a
level of service or another psychological parameter, we can only
indirectly measure this variable. Data from the social world is
often qualitative and requires different considerations than in the
physical world.

The complex relationship between the data and the reality it seeks
to improve emphasises the need for subject-matter expertise about
the problem under consideration. Data should never be seen as
merely an abstract series of numbers or a corpus of text and images,
but should always be interpreted in its context to do justice to the
reality it describes.
Good Data Science 24

2.2.2 Data

Data is the main ingredient of data science, but not all data sources
provide the same opportunities to create useful data products.
The quality and quantity of the data determine its value to the
organisation. This mechanism is just another way of stating the
classic Garbage-In-Garbage-Out (GIGO) principle. This principle
derives from the fact that computers blindly follow instructions,
irrespective of truthfulness, usefulness or ethical consequences of
the outcomes. An amazing algorithm with low quality or insuffi-
cient data will be unable to deliver value to an organisation. On the
other hand, great data with an invalid algorithm will also result in
garbage, instead of valuable information.
The quality of data relates to the measurement processes used to
collect the information and the relationship of this process to the
reality it describes. The quality of the data and the outcome of the
analysis is expressed in their validity and reliability. Section 2.3
discusses the soundness of data and information in more detail.
The next step is to decide how much data to collect. Determining
the appropriate amount of data is a balancing act between the cost
of collecting, storing and analysing the data versus the potential
usefulness of the outcome. In some instances collecting the required
data can be more costly than the benefits it provides.
The recent steep reduction in the cost of collecting and storing data
seems to render the need to be selective in data gathering a moot
point. Data scientists might claim that we should collect everything
because a time machine is much more expensive than collecting
and storing more data than strictly necessary. Not measuring parts
of a process has an opportunity cost because future benefits might
not be realised. This opportunity cost needs to be balanced with the
estimated cost of collecting and storing data.
This revolution in data gathering is mainly related to physical
measurements and the so-called Internet of Things, mobile phones
Good Data Science 25

and other wearable devices. Measurement devices and the trans-

mission and storage of data are affordable, so carpet-bombing a
region with sensors to collect data becomes more feasible. Some
aspects of reality remain complicated and expensive to measure
as the Internet of Things cannot be applied everywhere. Assessing
how people feel about something, their intentions and so on will,
until we have access to cost-effective mass mind reading, remain a
complicated and expensive undertaking.
One guideline to determine what and how often to collect is to
work backwards from the sought benefits. Following the knowl-
edge pyramid, we should collect data that enables us to influence
reality positively. The frequency of collection is an outcome of the
statistical power that is required to achieve the desired objectives.
In most cases, the more data is available, the higher the statistical
power of the analysis.
The amount of data points required to achieve a specific outcome
also depends on the type of data. The more reliable and valid the
measurements, the fewer data points are needed to obtain a reliable
sample. Lastly, the need to ensure that a sample represents the
population it describes defines the minimal size of the sample in a
social context. Determining sample sizes is a complex topic and the
statistics literature provides detailed information about how much
data to collect to achieve the required statistical power.
Gathering data about people because it might be useful in the future
also has ethical consequences. Storing large amounts of personal
data without a defined need can be considered unethical because
the data might be used for a purpose for which the subjects did not
consent. Medical records are a case in point. They are collected to
manage our health and not for insurance companies to maximise
their profits. The last section of this book discusses the ethics of
data science.
A case study illustrates how to decide the ideal amount of data to
collect. A water utility discussed how much data they wanted to
Good Data Science 26

gather to measure how much water customers use. The existing

method only provided one data point for each water meter every
three months. The water engineers would ideally like a reading
every five minutes, while the billing department was more than
happy with one daily reading.
New technology became available that collects data at a higher
frequency. However, the higher the rate, the higher the cost of
collection due to transmission bandwidth and battery life. Col-
lecting data every five minutes was considered to be unfeasible
and potentially unethical because it reveals too much about the
lifestyles of customers. Daily data was insufficient to provide
benefits in network design and operation. The utility decided to
collect hourly data because it allows for most of the sought benefits,
doesn’t significantly impact the privacy of customers and is within
reasonable reach for the current level of technology.

2.2.3 Information

Within the context of this book, information is defined as processed

data. Information is data placed with the context of the reality from
which it was extracted. To ensure information is sound and useful,
professionals need to use an appropriate methodology, logically
present the information and preserve the results for future reuse
or review.
Data scientists use an extensive range of methods to convert data
into information. At the lowest level, summarising the averages
of the various data points converts provides some value. More
complex analysis transforms data about the past into a prediction
of the future. These techniques require a solid understanding of
mathematics and analytical methods to ensure they don’t result in
data pseudo data science.
Communicating information is where art meets data science. Writ-
ing good reports and designing meaningful visualisations ensures
Good Data Science 27

that a data product is useful.

Lastly, the information needs to be preserved so that it is accessible
to those who need it in the future or for those who seek to review
the methods.
The last two sections in this chapter discuss the criteria for the
soundness of data products and how to best visualise information.
Chapter three delves into the various methods to convert data
into information. The last chapter covers some issues concerning
communicating the results of data science to other professionals.

2.2.4 Knowledge

Professionals with subject-matter expertise gain knowledge from

the results of data science, which they use to decide on future
courses of action. Knowledge can be either tacit or explicit. The
results of a data science project, or a data product, are explicit
knowledge. This knowledge can be transferred through writing or
instruction.
Numbers and visualisations help professionals to understand the
reality they need to manage. This process of understanding and
using these results in practice leads to tacit knowledge, which is the
essence of domain expertise. Tacit knowledge is difficult to transfer
because it consists of a combination of learnt explicit knowledge
mediated through practical experience.
Data science thus not only requires domain expertise to be useful,
but it can also increase this expertise. This topic is outside the scope
of data science as it ventures into knowledge management.

2.2.5 Feedback Loop

The last and most important part of this data science model is the
feedback loop from knowledge back to reality. The arrow signifies
Good Data Science 28

actionable intelligence which is how reality is improved through

knowledge. The key message of this section is that the results of
data science need to either lead to a different way of thinking about
a problem or provide actionable intelligence to propose.

Either option eventually leads to improved decisions using the best

available data and analysis. Care needs to be taken, however, that
the correct conclusions are drawn. The GIGO principle only covers
the input of the process, but also the process itself needs to be
sound. Although the data might be of good quality, a lousy analysis
will still result in ‘garbage’. The next two sections discuss how we
can ascertain whether the outcomes of data science are sound and
ensure the user draws the correct conclusion from the information.

2.3 Sound Data Science

Just like a building should be sound and not collapse, a data

product needs to be sound to be able to create business value.
Soundness is where the science and the data meet. The soundness
of a data product is defined by the validity and reliability of the
analysis, which are well-established scientific principles (Figure
6).³ Soundness of data science also requires that the results are
reproducible. Lastly, the data and the process of creating data
products need to be governed to assure beneficial outcomes.

The distinguishing difference between traditional forms of business

analysis and data science is the systematic approach to solving
problems. The key word in the term data science is thus not data
but science. Data science is only useful when the data answers a
useful question, which is the science part of the process.⁴
³Anderson, C. (2015). Creating a Data-Driven Organization: Practical Advice
from the Trenches. Sebastopol, CA: O’Reilly Media Inc.
⁴Caffo, B., Peng, R., & Leek, J.T. (2018). Executive Data Science. A Guide to
Training and Managing the Best Data Scientists. LeanPub.
Good Data Science 29

This systematic approach ensures that the outcomes of data science

can be relied upon to decide on alternative courses of action.
Systematic data science uses the principles of scientific enquiry,
but it is more pragmatic in its approach. While scientists search for
general truths to explain the world, data scientists pragmatically
seek to solve problems. The basic principles that underpin this
methodical approach are the validity, reliability and reproducibility
of the data, the methods and the results.

Figure 6: Visualising validity and reliability.

2.3.1 Validity

The validity of a data set and the information derived from it relates
to the extent to which the data matches the reality it describes. The
validity of data and information depends on how this information
was collected and how it was analysed.
For physical measurements, validity relates to the technology used
to measure the world and is determined by physics or chemistry. If,
for example, our variable of interest is temperature, we use the fact
that materials expand, or their electrical resistance increases, when
the temperature rises. Measuring length relies on comparing it with
a calibrated unit or the time it takes light to travel in a vacuum. Each
type of physical measurement uses a physical or chemical process,
and the laws of nature define validity. When measuring pH, for
Good Data Science 30

example, we want to be sure that we actually measure the power

of hydrogen ions and not some other chemical property.
Psychological processes, such as customer satisfaction or personal-
ity, are much harder to measure than physical properties. Although
a state of mind is merely a pattern of electrical and chemical activity
in the brain, no technology can directly measure it. Not much is
known about the relationship between the physical events in the
brain and our feelings, motivations and other psychological states.
Not all data about people has a validity problem. Observations that
relate directly to our behaviours, such as technology that tracks
our movements, or eye-tracking equipment to record our gaze, are
physical measurements. Demographic data is a direct measurement
of a social state. However, even seemingly simple aspects such as
gender can lead to significant complexity when trying to measure
it. What often seems a simple demographic variable can be quite
complicated to define.
Scientists use complex machinery to scan brains to discover how
our mind functions. Marketers regularly use this technology to
understand customers. Neuromarketing produces insights from
brain scans to fine-tune the design of products and marketing com-
munication to maximise the likelihood that a customer purchases
their offering.⁵ Brain scanning gives insight into the processes
inside our brain, but it is quite expensive and intrusive. Scanning
technologies are insightful, but not an efficient way to get to know
your customers.
In practice, social scientists and psychologists use psychometrics
to indirectly measure states of mind by using survey techniques.
We might ask a customer whether they agree or disagree with a
series of statements such as “The hotel room was comfortable”.
Most commonly these questions are measured using a Likert scale
with five or seven descriptors.
⁵Lindström, M. (2010). Buy.Ology: Truth and Lies About Why We Buy. New York:
Broadway Books.
Good Data Science 31

The basic principle of surveys to measure a state of mind is that

the mental processes we seek to understand causes the subject to
answer a survey question in a certain way. People who think that
the hotel room was very comfortable will fully agree with the
statement in the survey. The statistical analysis of such questions
seeks to reverse this causality to learn about the subject’s thoughts
and feelings.

The variables that we are interested in are latent variables because

they are hidden within the mind of the subject and only reveal
themselves indirectly through the survey answers. The validity of
psychometric measurements is a complex topic with many types of
validity, each with their specific purpose. A vast field of literature
describes how to measure and analyse latent variables.⁶

Another method to understand people is to use a Big Data ap-

proach. This technology does not rely on surveys or brain scans
from a sample of the population but uses our behaviour on social
media or measured through a wearable device as a proxy for our
psychology. The big data approach is entirely different from any of
the other techniques.

Brain scanning and psychometrics aim to understand future be-

haviour by indirectly measuring what we think, feel and believe.
One of the main problems with surveys is that our answers are
biased and perhaps not an accurate reflection of what we believe.
Brain scans are impressive but are still only a proxy for our internal
states of mind.

Big data methods measure our actual behaviour by recording what

we purchase, where we travel, what we search for and so on.
The second significant difference between big data and traditional
approaches is that the first two methods rely on a small sample of
the population, while big data approaches, by their very nature,
⁶DeVellis, R.F. (2011). Scale Development: Theory and Applications (3rd ed.).
SAGE Publications.
Good Data Science 32

include millions of subjects. The validity of this information is

very high because it measures actual behaviour instead of indirect
parameters.

2.3.2 Reliability

The reliability or accuracy of physical measurements depends on

the quality of the instrumentation used to obtain the data. Engi-
neers spend a lot of effort to assure the reliability of instrumentation
through maintenance and calibration programs. The instruments
need to be installed and maintained to the manufacturer’s spec-
ifications to ensure their ongoing accuracy. Quality assurance of
instrumentation is possibly the most costly aspect of collecting and
storing data about physical processes.

Several methods exist to test the reliability of psychological survey

data. One simple test is to check for respondents that provide the
same answer to all questions. The chances that somebody would
genuinely answer “Neither agree or disagree” to all questions is
negligible, and it is good practice to remove these people from
the sample. Researchers also use questions to trap fake responses.
The researcher can, for example, ask people whether they agree
or disagree with certain factual statements, such as: “You live in
Australia.” Any subject not wholly agreeing with this question
(assuming this is an Australian survey) should be removed from
the sample.

Brain scanning technology relies on small samples of the popula-

tion because of the cost of the technology. These methods are sensi-
tive to error, illustrated by an infamous example. Craig Bennett and
Abigail Baird placed a dead Atlantic salmon in an fMRI scanner and
were surprised to find brain activity. They published their results
to warn scientists of the risk of unreliable statistical methods. The
spoof scientific journal Annals of Improbable Research awarded
Good Data Science 33

Bennet and Baird with the igNobel prize. This annual prize awards
unusual and trivial results in science.⁷

The reliability of big data approaches is arguably very high be-

cause people provide this information not to satisfy the need of a
researcher but to guide their actions. Rather than asking somebody
whether they will purchase something in the future, our actual pur-
chase patterns are naturally a strong predictor for future purchases.

2.3.3 Reproducibility

The third aspect of the soundness of a data product is its repro-

ducibility, which is the ability for other people to reconstruct the
workflow of the analyst from raw data collection to reporting. This
requirement is a distinguishing factor between traditional business
analysis and data science. The condition of reproducibility ensures
that managers who base business decisions on a data product can
review how the results were obtained or at least have trust in the
results because they are potentially auditable. Reproducibility en-
sures that peers can evaluate all analysis and negates the problems
of black boxes.

Replication is the gold standard for science to ensure the quality

of previous results. Replication involves repeating the experiments,
which is a costly and complex activity. The past few years, some
areas of science have suffered from a replication crisis. Famous
experiments in psychology and sociology are under a cloud of
doubt because current researchers fail to replicate the findings.
Some scientists were clearly fraudulent in their approach, changing
data and manipulating the analytical process to obtain the desired
results. Other scientists either used unrepresentative samples or
⁷Bennett, C.M., Baird, A.A., Miller, M.B., & Wolford, G.L. (2009). Neural
correlates of interspecies perspective taking in the post-mortem Atlantic Salmon:
an argument for multiple comparisons correction. NeuroImage, 47, S125. DOI
10.1016/S1053-8119(09)71202-9.
Good Data Science 34

misinterpreted the data and published wrong results. The peer

review process did not identify these issues because the authors
only provided the outcomes of their research, and the papers were
published and accepted as truthful.
This crisis in some of the sciences has moved some journals to
require authors to submit not only the written text but also the
data and analytical code that was used to draw the conclusions.
Releasing the data and the code strengthens peer review. These
principles of reproducibility also apply to business analysis and
ensure sound data science. Every study needs to be undertaken in
such a way that a colleague or auditor can recreate the steps of the
recipe that was followed to reach a conclusion.
The most effective method to achieve full reproducibility is to
use literate programming. This method combines computer code
with text so that the analysis becomes fully transparent. Although
many systems exist that at first instance might seem more user-
friendly than writing code, point-and-click systems have severe
limitations, and the results are often difficult to verify. The last
chapter discusses the various tools to create data products in more
detail.
A specific aspect of machine learning that relates to its repro-
ducibility is whether the user can understand how the model drew
a conclusion. Can the data scientist explain to the subject-matter
expert how the model works? The great benefit of machine learning
is that these algorithms can detect patterns in large volumes of data
that are impossible for humans to comprehend within a lifetime.
More often than not, however, machine learning results in a black
box that converts data into output. Although the algorithms are
reproducible in that they produce the same result with the same
data, they are not necessarily explainable. The outcomes of, for
example, a random forest model, can take tens of pages to print
and the logic will be hard to verify for humans.
Machine learning algorithms are sometimes simplified so that the
Good Data Science 35

responsible managers can understand the logic of the decisions they

make. In these cases, reliability is sacrificed for greater explainabil-
ity. Whether a data science product is explainable depends not only
on the code itself but also on the level of mathematical insight of the
consumer of the outcomes. Explainability is thus a direct result of
the level of data literacy of the organisation. This theme is further
discussed in the final chapter.

2.3.4 Governance

The fourth aspect of sound data science is governance. The process

of creating data products needs to be documented in line with qual-
ity assurance principles. Practical considerations such as naming
conventions for scripts, coding standards to ensure readability and
so on are a necessary evil when managing complex data science
projects.

The same principle also applies to managing data. Each data source
in an organisation needs to have an owner and a custodian who
understands the relationship between this data and the reality from
which it is extracted. Large organisations have formal processes
that ensure each data set is governed and that all employees use a
single source of truth to safeguard the soundness of data products.

Governance is a double-edged sword as it can become the ‘wet

blanket’ of an organisation. When governance becomes too strict,
it will smother the very innovation that data science is expected
to deliver. The art of managing a data science team is to find a
middle way between strictly following the process and allowing
for deviations of the norm to foster innovation. Good governance
minimises risk, while at the same time enabling positive deviance
that leads to better outcomes.
Good Data Science 36

2.4 Aesthetic Data Science

Vitruvius insisted that buildings, or any other structure, should

be beautiful. The aesthetics of a building causes more than just
a pleasant feeling. Architecturally designed places stimulate our
thinking, increase our well-being, improve our productivity and
stimulate creativity.

While it is evident that buildings need to be pleasing to the eye,

the aesthetics of data products might not be so obvious. The
requirement for aesthetic data science is not a call for beautification
and obfuscation of the ugly details of the results. The process of
cleaning and analysing data is inherently complex. Presenting the
results of this process is a form of storytelling that reduces this
complexity to ensure that a data product is understandable.

The data science value chain starts with reality, described by

data. This data is converted to knowledge, which managers use to
influence reality to meet their objectives. This chain from reality
to human knowledge contains four transformations, each with
opportunities for a loss of validity and reliability. The last step
in the value chain requires the user of the results of data science
to interpret the information to draw the correct conclusion about
their future course of action. Reproducibility is one of the tools
to minimise the chance of misinterpretation of analyses. Another
mechanism to ensure proper interpretation is to produce aesthetic
data science.

Aesthetic data science is about creating a data product, which can

be a visualisation or a report, that is designed so that the user draws
correct conclusions. A messy graph or an incomprehensible report
limits the value that can be extracted from the information. The
remainder of this section provides some guidelines on designing
good visualisations and writing reports.
Good Data Science 37

2.4.1 Visualisation

Data visualisations are everywhere. They are no longer the domain

of scientific publications and business reports. Publications in every
medium use graphs to tell stories. The internet is awash with
infographics on a wide range of topics. These popular images are
often data science porn because they are designed to entertain and
titillate, with limited usability from a business perspective. They
are a fantastic tool to supply information to customers but should
not be used to report data science.
Aesthetics and usefulness go hand in hand. Some data visualisa-
tions in engineering remind me of a Jackson Pollock painting, with
multitudes of lines and colours splashed over the screen. Adding too
much information to a graph and using too many colours reduces
its usability. When visualisation is not aesthetic, it becomes harder
to interpret which leads to the wrong conclusions and can even
deceive the user.⁸
The previous chapter discussed how Taylor introduced the concept
of scientific management to remove bias from business decisions. A
data scientist needs to be aware of these biases in order to prevent
them and create data products that don’t deceive. Many of these
biases relate to how information is presented.
Our perception is not always an accurate representation of reality,
and we often misinterpret the images that our retina collects.
Optical illusions are funny internet memes, but they also occur in
real life. Besides optical illusions, messy visualisations are hard to
interpret because our mind does not know what element to focus
on. A messy graphic confuses the brain so that it starts to form its
interpretations.
Perhaps a good data visualisation should look more like a painting
by Piet Mondrian who is famous for his austere compositions with
⁸Jones, G. E. (2007). How to Lie with Charts (2nd ed). Santa Monica, Calif:
LaPuerta.
Good Data Science 38

straight lines and primary colours. Using art to explain data visual-
isation is not an accidental metaphor because visual art represents
how the artist perceives reality. This comparison between Pollock
and Mondrian is not a judgement of their artistic abilities. For
Pollock, reality was chaotic and messy, while Mondrian saw a
geometric order behind the perceived world.
Although visualising data has some parallels with art, it is very
different. All art is basically a form of deception. The artist paints
a three-dimensional image on a flat canvas, and although we see
people, we are just looking at blobs of paint. Data visualisation as an
art form needs to be truthful and not deceive, either intentionally
or accidentally. The purpose of any visualisation is to validly and
reliably reflect reality.
Aesthetic data science is not so much an art as it is a craft. Following
some basic rules will prevent confusing the consumers of data
products. Firstly, a visualisation needs to have a straightforward
narrative. Secondly, visualising data should be as simple as possible,
minimising elements that don’t add to the story.

2.4.1.1 Storytelling

First and foremost, a visualisation needs to tell a story. The story

in data visualisation should not be a mystery novel or of the
choose your own adventure genre. A visualisation should not have
suspense but get straight to the point. Trying to squeeze too much
information into one graph will confuse the reader. Ideally, each
visualisation should contain only one or two narratives. If there is
more to tell, then use more charts and create a dashboard.
Numerical data can contain several types of narratives. A graph can
compare data points to show a trend among items or communicate
differences between them. Bar charts are the best option to compare
data points with each other. A line graph is possibly your best
option to compare data points over time. The distribution of data
Good Data Science 39

points is best visualised using a histogram. Scatter plots or bubble

charts show relationships between two or three variables (Figure
7).

Figure 7: Examples of stories with quantitative data

The detailed considerations of choosing the most suitable visualisa-

tion are outside the scope of this book. The Chart Chooser website
(chartchooser.com) provides a dynamic interface to choose the best
graph to tell a story. The main point is that every visualisation
needs to tell a story and not just summarise a bunch of data.
Visualising qualitative information has its own language with
many options to tell a story. Displaying qualitative information is
more an art than a craft because there is less reliance on mathe-
matics. Word clouds are a popular tool to summarise text. These
visualisations display the relative frequency of the words in a text
(Figure 8). Network diagrams are another common visualisation
tool. Networks are a convenient method to analyse relationships
between people or other qualitative entities such as journal articles.
Good Data Science 40

Figure 8: Wordcloud of a draft version of this chapter with top-25 words.

2.4.1.2 Visualisation Design

Beauty is in the eye of the beholder, and there are no formulas

or algorithms to ensure perfect visualisations. The social network
Reddit has two groups dedicated to visualisations. Users members
of the Data is Ugly and Data is Beautiful groups share images of
visualisations they consider ugly or beautiful. These two groups
sometimes share the same visualisations because of different inter-
pretations of aesthetics in data. What is a beautiful visualisation to
one person, is an abomination to somebody else. The aesthetics of
data visualisation is for a significant part in the eye of the beholder.
However, when viewing aesthetics from a practical perspective, we
can define what this means with a simple heuristic.

Edward Tufte is an American statistician who is famous for his

work on visualisation.⁹ Tufte introduced the concept of the data-
ink ratio. In simple terms, this ratio expresses the relationship
between the ink on the paper that tells a story and the total amount
of ink on the paper. Tufte argues that this ratio should be as
close to one as possible. In other words, we should not use any
⁹Tufte, E. R. (1997). Visual Explanations: Images and Quantities, Evidence and
Narrative. Cheshire, Conn.: Graphics Press.
Good Data Science 41

graphical elements that don’t communicate any information, such

as background images, superfluous lines and text.

Now that we are in the paperless era, we can use the data-pixel
ratio as a generic measure for the aesthetics of visualisations.
The principle is the same as in the analogue days. Remove any
redundant information in your visualisation. Unnecessary lines,
multiple colours or multiple narratives risk confusing the user of
the report.

The data-ink ratio is not a mathematical concept that needs to be

expressed in exact numbers. This ratio is a guideline for designers
of visualisations to help them decide what to include and, more
importantly, what to exclude from an image.

Figure 9 shows an example of maximising the data-ink ratio. The

bar chart on the left has a meagre data-pixel ratio. The background
image of a cat might be cute and possibly even related to the topic
of the visualisation, but it only distracts from the message. Using
colours to identify the variables is unnecessary because the labels
are at the bottom of the graph. The legend is not very functional
because it also duplicates the labels. Lastly, the lines around the
bars have no function.

To improve this version, all unnecessary graphical elements have

been removed. Assuming that the story of this graph is to compare
variables, the columns have been ranked from large to small. If the
narrative of this graph was to compare one or more of the variables
with other variables, then groups of bars can be coloured to indicate
the categories.

The basic rule of visually communicating data is to not ‘pimp’

your visualisations with unnecessary graphical elements or text
that does not add to the story. When visualising data, austerity is
best practice.
Good Data Science 42

Figure 9: Examples of the data-pixel ratio.

2.4.1.3 A Splash of Colour

In the era before the petrochemical revolution, colours were mostly

limited to nature because artificial colouring was very expensive.
The purple togas of Roman emperors were dyed with the secretion
of thousands of sea snails. Seventeenth-century Dutch painter
Johannes Vermeer used precious lapis lazuli to obtain the perfect
blue for his paintings. The expense of creating colours in the past
limited the art of choosing the right ones to artists and rich people.

In our times, we have easy access to every imaginable colour. We

can buy clothes in whatever colour or desire or paint our house
like Pippi Longstocking’s villa. We are spoilt for choice concerning
colour. The increased possibilities are also a burden because we
need to choose the right colour, not just any colour. This tyranny
of choice also plays a role in designing beautiful visualisations.

When buying clothing or painting a house, our choice of colours is

hopefully governed by a sense of style. When designing visualisa-
tions, colour plays a more instrumental role. The colours in a graph
are not decoration, but they communicate a narrative. The basic
Good Data Science 43

principle for good visualisation design is to minimise the number

of colours.
Colours have intrinsic meaning that helps to tell the story. Red
is generally recognised as a warning sign and green a calmer
tone. In Western cultures, pink associates with femininity and the
colour blue is often used in a corporate sense. Marketers use these
associations with colour when designing brands. The psychological
meaning of colour is, however, not culturally constant and can
vary between countries. The Color Psychology website (colorpsy-
chology.org) contains valuable information about how colours are
interpreted.
Traffic lights are one of the most popular colour schemes in business
reporting. The colour red stands for an adverse outcome, yellow
for a neutral one and green is positive. This type of reporting
helps managers to focus on problem areas so they can discuss
actions to improve future performance. A note of caution is that this
technique does not work for men with green/red colour blindness.
This condition is not a problem with real traffic lights as the order
of the lights is always the same. However, on a business report, the
colours will all look the same to roughly eight per cent of men with
this condition.
The minimum amount of colour depends on the narrative you
want to convey and the type of visualisation. Cartography often
uses colour to communicate its message. One of the problems that
plagued mathematicians for centuries is how many colours we
need as a minimum to colour every area on a map, without two
bordering regions with the same colour. Cartographers know from
practical experience that this number is four, but mathematical
proof only came in a few decades ago.
Besides minimising the number of colours, we also need to know
which colours to use. Cartographers Mark Harrower and Cynthia
Brewer developed the Color Brewer system (colorbrewer2.org) to
help designers of visualisations select a good scheme. These colour
Good Data Science 44

schemes are designed for choropleth maps, but can also be used for
non-spatial visualisations. ¹⁰ The Color Brewer system consists of
three types of colour palates: sequential, diverging and qualitative
(Figure 10).

Figure 10: Types of colour schemes.

Sequential schemes contain a series of colours with increasing

strength. These colour schemes are most suitable to visualise mag-
nitude from low to high, with light colours usually for low data
values and dark colours for high values.
Diverging colours are used to visualise a deviation from a norm,
such as droughts or floods or adherence to a budget. Green, yellow
and red are the most common use of this type of palate as business
reports are filled with traffic lights to report progress.
Qualitative colours are groups of colours that are aesthetically
compatible but without a logical relationship between them based
on the data. These palates can express qualitative values such as
categories. The left graph in figure 9 is an example of a qualitative
scheme to indicate categorical variables.

2.4.2 Reports

Advertising executive Fred Barnard coined the now famous idiom

that “a picture is worth (ten) thousand words” in 1927. While this
¹⁰Harrower, M., & Brewer, C.A. (2003). ColorBrewer.org: An Online Tool for
Selecting Colour Schemes for Maps. The Cartographic Journal, 40(1), 27–37. DOI
10.1179/000870403235002042.
Good Data Science 45

might be the case, the complexity of data science in most cases

requires text to explain the analysis.

To claim that a report needs to be written with clarity and precision

in proper spelling and grammar almost seems redundant. The
importance of readable reports implies that the essential language
a data scientist needs to master is not Python or R but English, or
another relevant human language.

Writing a good data report enhances the reproducibility of the

process by describing all the steps in the process. A report should
also help to explain any complex analysis to the user to engender
trust in the results.

The topic of writing good business reports is too broad to do justice

within the narrow scope of this book. For those people that need
help with their writing, data science can also assist. There are
many great online writing tools to support authors not only with
spelling but also grammar and idiom. These advanced spelling
and grammar checkers use advanced text analysis tools to detect
more than spelling mistakes and can help fine-tune a text using
data science. As English is my second language, I rely heavily on
the Grammarly software to ensure it is free of apparent issues.
However, even grammar checking with machine learning is not
a perfect replacement for a human being who understands the
meaning of the text.

2.5 Best-Practice Data Science

This chapter discusses a normative framework for creating value

with data. Inspired by the Roman architect Vitruvius, the products
of data science need to be useful, sound and aesthetic.

Data science is useful when the data is converted to information.

This information increases the knowledge of the professionals who
Good Data Science 46

use it. This knowledge improves the reality from which the data
was extracted. The relationship between reality and data is critical.

Data science is sound when it delivers valid and reliable results and
can be reviewed by other experts.

Validity is the extent to which the data describes the aspect of

reality it is presumed to represent. In physical measurements, this
aspect is governed by physics and chemistry. In measures of the
social world, validity is complicated because we can only record
the external states of a human being and not their state of mind.

The reliability of data and the analysis relates to the accuracy of

the information. In physical measurements, accuracy is a function
of how well an instrument is installed and maintained. Within the
social sciences, reliability relates closely to the biases of people
providing data.

Sound data science also needs to be reproducible and well governed.

Reproducibility is essential to enable other people to understand the
steps taken to achieve a result. Governance is a crucial aspect of
managing a business and helps with making data science efficient
and effective.

Lastly, the results of data science need to be aesthetic to reduce

the likelihood that users draw the wrong conclusions. This need
for aesthetics does not imply beautification but a focus on telling
a story with the data. The basic principle of good visualisations
is to minimise text and graphics that does not communicate any
information.

The next chapter delves into the strategic aspects of data science by
presenting a five-phase framework that organisations can follow to
enhance the value they extract from data.
3. Strategic Data Science
Managers often claim to be “data rich but information poor”. This
statement is in many cases only partially correct because it hides a
misconception about the life cycle of data. Being replete with data
but poor concerning information suggests that previously untapped
data sources are waiting to be mined and used.
It is highly unlikely that any organisation collects data for no partic-
ular purpose. Data is in most cases collected to manage operational
processes. Collecting data without purpose is a waste of resources.
After the data is used, it is stored and becomes ‘dark data’. Because
almost all business processes are recorded electronically, data is
now everywhere. Managers rightly ask themselves what to do with
this information after it is archived. A strategic approach to data
science helps an organisation to unlock the unrealised value of these
stores of data to better understand their strategic and operational
context.
Whereas the framework in the previous chapter is normative and
defines ideal data science, the model in this chapter describes
the path that an organisation can take to increase the value they
extract from data. The data science continuum is a hierarchical five-
phase framework towards becoming a data-driven organisation.
This chapter discusses the phases of this continuum as a strategy
map for data science.

3.1 The Data Science Continuum

Strategic data science follows a process where an organisation

builds competencies towards high-value analysis. The evolution
Strategic Data Science 48

of becoming a data-driven organisation starts with collecting data

generated during operational processes. The next step involves
describing this data using exploratory techniques such as basic
visualisations and summary statistics, which is the domain of
traditional business reporting.

Figure 11: Data science continuum.

After the data is explored and understood, organisations can diag-

nose business processes to understand causal and logical relation-
ships between variables. The penultimate phase in the continuum is
to use knowledge of the past and its causal and logical connections
to predict possible futures and construct the desired future. The
final stage in the data science journey is a situation where data is
used to prescribe day-to-day operations (Figure 11).
The continuum is not a journey from left to right, to eventually land
at a place where algorithms control our destiny, and the remainder
of the continuum becomes less critical. The continuum guides data
science strategy towards this point but forms a strict hierarchy.
Before algorithms can independently decide anything, you need to
be able to predict the immediate future. To predict the future, you
need to have a good grasp of the descriptive statistics to diagnose a
business process. Lastly, the GIGO principle demands that analysis
is only possible if we understand the collected data.
Strategic Data Science 49

3.2 Collect data

The ultimate purpose of collecting data is to enable decisions that

improve reality or reduce the risk of future adverse impacts. The
previous chapter discussed the importance of understanding the
relationship between data and the reality it describes. This rela-
tionship is the essence of domain knowledge. Professionals in their
respective domains are trained and experienced with measuring the
reality they manage. Chapter two described some of the differences
in measurement between the two domains.

The literature often distinguishes between raw data and processed

data as the main ingredients of analysis. The idea that data can
be raw and natural is deceiving. There is no such thing as raw
data because every time we collect information from a physical
or social process, we need to decide how the data is collected.
These decisions are always informed by assumptions about what
this reality looks like before we see the data. We cannot measure
anything if we don’t know what it is we like to know.

In addition to these assumptions, we also need to choose which

aspects of a process to measure. Every data set is only a sample of
reality informed by our assumptions. No amount of information
will ever adequately describe the physical or social reality that we
want to manage. All data collection is a process of compromise and
assumptions. Understanding these assumptions and limitations is
an essential step towards creating value.

The importance of the quality of the data cannot be understated.

The Garbage-In-Garbage-Out (GIGO) principle is inescapable, even
when using the most sophisticated machine learning algorithms.
Understanding the quality of data requires subject-matter exper-
tise and mathematical skills. The GIGO principle underwrites the
importance of subject-matter experts who understand the flow of
data from its creation to the point of analysis.
Strategic Data Science 50

Data is always collected and stored to manage a business process.

When a customer calls to lodge a complaint, the operator records
this information to facilitate the customer’s experience. When a
treatment plant measures tank levels, the control system uses this
information to decide when to run the filters to ensure the town
does not run out of water. All data is used at or near the time of
its creation. After the process has completed, it becomes dark data.
This dramatic name does not do justice to the data and makes it
seem like servers are full of useless bits and bytes. The data was,
however, useful when it was generated and is merely waiting to be
recycled.

Much of data science involves recycling this information to create

additional value. In the case of a customer service process, we might
analyse whether complaints cluster within a particular region. For
the treatment plant, we can use this data to assess the performance
of the plant to help us optimise energy consumption. Since the
data revolution, more data is generated, used and archived than
ever before. One of the powers of data science is to reuse this
resource to gain additional insight to improve performance. A
data scientist or business analyst is like an archaeologist who digs
through electronic archives to make sense of the past to influence
the future.

The salient point is that data is rarely collected for the purpose
it was initially intended. This means that all data needs to be
transformed into a format that is suitable to achieve the objectives
of the project. Cleaning data is the least exciting activity of data
science, but it often requires much of the available resources. The
informal terms used in the industry for cleaning data are wrangling,
munching or data jujitsu. These not so positive references illustrate
the effort and frustration associated with cleaning data.

Reality is fickle which means that we never have access to all

possible data about our variable of interest. People filling out a
survey might not want to answer all questions due to privacy fears,
Strategic Data Science 51

survey fatigue or short attention spans. Technology to collect data

is never perfect, and also physical measurements are often missing
for various reasons.

Missing data needs to be managed by either imputing the gaps

with estimated values or by removing parts of the sample. When
only a few data points are missing, statistical methods can fill the
gaps using various methods. Simple techniques, such as replacing
missing data with the previous value or the mean, to more complex
procedures that use machine learning, minimise the amount of
missing data. When large chunks of data in surveys are missing,
subjects with a low percentage of responses should be removed
from the sample. Preventing the GIGO syndrome means that the
data scientist first needs to remove the garbage from the data set.
The most suitable method to manage missing data depends on the
reasons it was missing and the impact of the missing points.

The data revolution has spawned the phenomenon of big data,

briefly mentioned in section 2.3. The issues regarding big data
are mainly technical in nature concerning storing, managing and
analysing this data. Accessing and analysing billions of records
requires specialised techniques, not relevant in small sets of data.

Besides its volume, big data is characterised by three other prop-

erties starting with the letter V. Big data has a lot of variety
compared to traditional sources. The databases of quintessential big
data organisations like Google and Facebook contain information
about a wide range of behaviours of its users. This data provides
a digital fingerprint of our online lives, from the music we play to
our network of friends.

Thirdly, big data has a very high velocity, resulting in many

measurement points in a short period. This high velocity contrasts
with the traditional sampling methods that only collect information
every so many minutes, hours or even days.

Lastly, big data has a high level of veracity. Traditional data sources
Strategic Data Science 52

about human behaviour ask people what they feel or what they
might do in the future. This approach has a low level of veracity
because they are indirect measurements of behaviour. Big data has
a high level of reliability because it measures actual conduct.

The combination of this high volume, great variety and velocity

and veracity gives big data its enormous power. At no point in
human history have we had access to the detailed information
about our behavioural patterns. The vast majority of data sets used
in organisations are, however, not big data.

Big data provides previously impossible opportunities for organi-

sations that can generate this type of information. A data set does,
however, not have to be big to create business value. Whether a
set of data falls into the category of big data is not a concern
from a strategic perspective but merely a technical problem. The
main point in collecting evidence is to ensure that the information
helps the organisation to make evidence-based decisions. The same
principles apply to all sources of data, big and small.

3.3 Descriptive Statistics

The first step in any analytical process is to describe the data under
consideration. Descriptive statistics are methods, such as the aver-
age, median, range, standard deviation, and so on, that summarise
a data set. Business intelligence and exploratory analysis are the
most common uses of descriptive statistics.

Descriptive statistics are the rear-view mirror of management. Most

performance reports consist of large volumes of tables and graphs
that report on what happened in the previous reporting period.
Business reporting is often viewed as a lowly aspect of data science
because you need to look through the windshield to drive a car,
and not just look through the rear-view mirror. More crudely, some
Strategic Data Science 53

say that traditional performance reporting is like counting the dead

after the battle has been fought.
Although this analogy might seem a compelling argument to ditch
the traditional business report, it does not negate the importance
of a rear-view mirror. The car analogy is also logically invalid
because what we see through the windscreen is not the future, but
the present. Business processes use data from the present through
monitoring, and performance reporting teaches us about the past
to learn to create a better future. Descriptive statistics are not a
sufficient condition for becoming a data-driven organisation but
are nevertheless a necessary condition. Descriptive statistics are the
unavoidable first step towards more complex analysis.

3.3.1 Business Reporting

When developing a business report, managers need to make two

decisions. Firstly, they decide which performance indicators are
important enough to include in the business report. The second
decision is how to present this information so that readers of
the report draw valid conclusions. Both choices require balancing
various considerations. Choosing which performance indicators to
include in a performance report is a fine art that needs to balance
the requirement for completeness and consistency. Developing a
performance dashboard is an iterative process where the measures
are continuously fine-tuned.
The famous words of former General Electric CEO Jack Welch are
frequently touted in board rooms around the globe: “What you
measure is what you get”.¹ This phrase is used so often that it has the
status of a business axiom, the truth of which cannot be doubted.
This rule suggests that the more you measure, the better you can
control the destiny of your organisation. Welch seems to warn us
that any process missing from the business report is doomed to fail.
¹Welch, J. (2001). Straight from the Gut. Warmer Books.
Strategic Data Science 54

These are wise words, but measuring too many variables often leads
to confusion instead of enlightenment. It seems that the managers
that use these words did not read the whole book. Welch also
moderates his insistence on measurement and writes that: “Too
often we measure everything and understand nothing”. It seems
that business reporting hides a paradox.

No matter how many KPIs or other parameters you measure, a

performance measurement system will never be both complete and
consistent. This means that adding more and more traffic lights,
graphs, maps and other elements to a dashboard increases the risk
of confusion and inconsistency. The reverse side of this problem is
that a minimalist dashboard of consistent descriptive statistics will
be incomplete because not everything of interest is measured.²

Once an analyst selects the ideal set of performance indicators,

they decide how to communicate these to management. Many
methods and tools enable reporting descriptive statistics. At the
lowest level, we can display the information in a table so that users
can explore the data for themselves. While this method provides
excellent transparency, it leaves users to develop their own story
about the information.

Conditional formatting to highlight some of the salient information

enhances the tables, but is not an ideal way to communicate data.
Data visualisation, as discussed in the previous chapter, is a much
more effective method to tell a story with data than tables. Best
practice is to show the visualisation and provide access to the
graphed data in a table to promote full reproducibility and increase
trust.
²Prevos, P. (2012). The incompleteness theorem of performance measurement
in service delivery. In World Business Capability Congress. Auckland: Centre for
Organisational Excellence Research (COER).
Strategic Data Science 55

3.4 Diagnostics

The vast majority of analytical projects involves diagnosis, which

is the process of finding causal or logical relationships between
variables. Analysing data uses mathematical transformations to
find and validate relationships between variables. We might need
to know whether complaints are clustered in a certain region of a
service area, or find the most likely cause of those complaints by
relating it to other operational data.

This description implies that visualising data to show a distribution

or compare two or more variables is strictly speaking not analysing
anything. Descriptive statistics in most performance reports re-
duces the data to fewer numbers, but strictly speaking, does not add
any information to the data set. The average of a set of numbers or
a trend line is already within the numbers. The defining property
of diagnosis or analysis is that the cleaned data is transformed to
reveal new information. An analysis shows us something that is
not apparent from the data itself by combining and processing the
information.

The analytical toolbox to find causal and logical relationships

is vast and contains many different approaches to creating new
knowledge. These methods cover a multitude of numerical tech-
niques. Regression helps to find simple patterns, principal com-
ponent analysis simplifies the data or Monte Carlo simulations
to manage uncertainty. To comprehensively discuss the available
methods is outside the scope of this book as the literature on this
topic is extensive.

Each domain of knowledge has its own specific methods to anal-

yse information. Engineers often use calculus and linear algebra
to solve problems. Marketers can use techniques such a Partial-
Least Square modelling to measure latent variables to understand
what their customers need and want. Economists have developed
Strategic Data Science 56

specialised econometrics to manage financial resources. Every field

of human endeavour has developed their own mathematical per-
spective of the world.

3.4.1 Qualitative Data Science

Most analysis centres on numerical or quantitative methods, but

there is also value in analysing qualitative information. Qualitative
analysis is often seen as being less valuable than working with
numbers because it is considered subjective. This subjectivity is
not a value judgement of qualitative methods but an inherent
issue in social science. Social scientists have developed methods to
manage this subjectivity by combining qualitative and quantitative
methods of analysis.

Computer-Assisted Qualitative Data Analysis is an emerging field

of computing that helps researchers to manage large amounts of
qualitative information using methods that minimise bias. These
software packages help to manage transcriptions of interviews or
focus groups, open-ended survey questions, social media or other
forms of text. These types of analysis require the researcher first
to untangle the texts and assign labels, codes or categories to
them. Once the data is processed by the researcher, the computer
assists in analysing the relationships numerically. Techniques such
as Grounded Theory or content analysis provide a systematic
approach to extracting value from qualitative data.

Other qualitative techniques involve the automated analysis of

texts and images. Sentiment analysis is a useful technique to assess
whether the sentiment of a section of text is positive, neutral or neg-
ative. Most of these methods rely on dictionaries of pre-classified
words. The algorithm counts their occurrence, which determines
whether the sentiment is positive or negative. For example, the
word “great” would count towards a positive sentiment, while the
word “cheesy” expresses a negative sentiment. Main problem with
Strategic Data Science 57

this technique is that many words in the English language are

homographs, which are words that are spelled the same way, but
have a different meaning. The word “lead” in a water context is
usually negative but the word “lead” in a business context could
express a positive sentiment.
Other automated text analysis techniques use numerical methods
to look for correlations between words. These text analysis methods
are useful to make sense of a large amount of text without having to
read each word. These methods reduce bias in analysing qualitative
information, but the output of such methods still needs to be
interpreted by the researcher. None of these methods can, however,
extract meaning from a text in the same way as a human would.

3.5 Predicting the Future

Predicting the future has been one of humanity’s primary quests

for millennia. Forces beyond our direct control influence our lives,
and predictions allow us to control the ever-changing world around
us to minimise risks to our lives. Before the age of enlightenment,
humanity tried to grasp the unpredictability of existence with
divination. They studied the movement of stars, animal entrails,
tea leaves and many other things to learn about their fate.
Since the scientific revolution, we no longer look at the stars, study
tea leaves or disembowel animals. Quantitative science is the most
successful form of divination ever invented. Many methods are
available to predict the future, from a simple linear regression
model to more complex random forests and neural networks are
available to connect the past with the future.
Organisations that seek to reap the rewards of data science are
often motivated by the capabilities of machine learning and other
advanced techniques. Being able to predict possible futures allows
a manager to assess whether this future is desirable. If this is not
Strategic Data Science 58

the case, then managers adjust current practice to create the desired
future. If a retail store can predict whether consumers are likely to
purchase a particular product because of their lifestyle preferences,
then they can target them in their marketing communication to
rattle their cage. If an engineer can predict that a piece of equipment
is likely to fail soon, then it can be replaced to avoid problems.

Predictive analysis uses information about the past and uses logic
to determine a possible future state of the variables in the model.
A water utility can, for example, develop a predictive model to
estimate the amount of water that their customers use on a given
day using their property size, affluence, weather and so on. A
retailer could develop a predictive model of the amount of store
traffic as a function of advertising, public holidays, sales, location
and whatever else might be predictive of this behaviour.

All reliable methods to predict the future use information from the
past, combined with a specified logic to estimate what could happen
next. This future only eventuates when all boundary conditions
under which this prediction was cast remain the same. No mathe-
matical model can predict the future, it can only provide insight
into possible futures. The purpose of predicting these possible
futures is to determine the best cause of action to create the desired
future.

This section provides a brief overview of predictive methods. The

book Predictive Analysis by Eric Siegel provides a comprehensive
introduction into methods to predict the future.³ The various meth-
ods to describe plausible futures consist of two approaches, each
based on different assumptions. Traditional methods follow the
scientific process of experimentally verifying hypotheses. Machine
learning is entirely different in that it does not use a hypothesis.
³Siegel, E. (2016). Predictive Analytics: The Power to Predict Who Will Click, Buy,
Lie, or Die. Wiley.
Strategic Data Science 59

3.5.1 Traditional Predictive Methods

The scientific method is the traditional process to develop a pre-

dictive model of the world (figure 12). This is a circular process
that describes an ongoing improvement in our understanding of
the world.⁴

A manager observes the process they are interested in through

descriptive statistics. Any anomalies that appear in the latest cor-
porate dashboard lead to new questions about what caused the
deviation from the desired outcome.

Figure 12: The process of science

The traditional method to answer these questions starts with a

hypothesis, which is a statement that can be confirmed or rejected.
A manager might want to know why sales are trending downwards
and hypothesise that a series of scorching days demotivated con-
sumers to visit the outdoor shopping centre caused the decline.
The analyst then develops a method to either verify or reject the
hypothesis, which increases our knowledge about the world. The
store manager can then use this knowledge to decide whether it
⁴Lee, N., & Lings, I. (2008). Doing Business Research: A Guide to Theory and
Practice. London: SAGE.
Strategic Data Science 60

is worth moving to an air-conditioned shopping mall to maximise

store traffic.

Some scientific processes lead to generalised knowledge of the

world in the form of a theory. The scientific view of a theory
is entirely different from the common sense understanding of
something that might or might not be true. A theory is a confirmed
relationship between variables measured in reality. For example,
the theory of gravity explains the forces between two objects
reliably predicts the motion of objects.

Scientists have developed thousands of theories to help us explain

the world. Scientific theories are a convenient way to predict and
thus control the future. The vast majority of theories that predict
the future are about the physical world. There are almost no
traditional theories in the scientific sense that allow us to predict
human behaviour, either individually or collectively reliably.

Once these theories are accepted and used, further observations will
raise new questions, and the same process is followed again. The
critical aspect to realise is that the insight of the researcher guides
this process. Formulating a reasonable hypothesis requires a deep
understanding of the subject matter under consideration.

The outcome of the scientific process is in many cases a formula

that describes the relationship between different states of reality.
Scientists prefer elegant theories. An elegant theory can be written
down in a relatively simple formula. The famous physicist Stephen
Hawking wrote in his inspiring book A Brief History of Time that
his dream was to one day have a formula that describes everything
in the universe that fits on a t-shirt.

3.5.2 Machine Learning

Machine learning is without a doubt the poster child of data science.

This popular technique possibly promises more than it is able to
Strategic Data Science 61

deliver. Machines cannot actually learn anything. Machine learning

is a group of complex algorithms that convert the input into output
by recognising patterns in data. The term was coined by Arthur
Samuel from IBM in 1959 to promote their capabilities in software
development.⁵

Machine learning is a branch of Artificial Intelligence (AI). The set

of algorithms that are classified as machine learning are part of the
AI toolkit but are not the only ways to mimic human reasoning.

Machine learning works differently to traditional prediction meth-

ods because it is the machines themselves that create the equations,
not the researcher. The family tree of machine learning has three
main branches: supervised and unsupervised methods and rein-
forcement learning (figure 13).

Figure 13: Overview of machine learning methods

In supervised methods, the algorithm is provided with a set of train-

ing data with known relationships. When, for example, wanting to
predict how much a viewer will like a particular movie, the training
data consists of verified information about what movies people
have watched and how they rated them. A supervised algorithm
⁵Burkov, A. (2019). The Hundred-Page Machine Learning Book. LeanPub.
Strategic Data Science 62

analyses the training data to discover patterns so that we can

predict how likely a viewer will enjoy a new movie. Supervised
learning can classify data in groups, or it can regress data to find
relationships between variables.

Recognising cats in images is an example of classification. The

algorithm is presented with thousands of labelled images, some of
which contain a cat and some don’t. The algorithm is taught how to
recognise a cat using these examples. Supervised learning is similar
to how humans learn to classify the world. As children, we are
provided with examples of cats by our parents, after we acquired
the skill, we independently classify furry animals as either a cat or
something else.

Regression analysis finds relationships between numerical or cate-

gorical variables using independent variables and known depen-
dent variables. The independent variables are the causes of the
dependent variable. We might know that the likelihood of rain
relates to temperature and moisture content of the air. This math-
ematical relationship is then used to predict the weather, estimate
life expectancy of people or assets and anything else that occurs
over time.

One of the most important considerations of supervised machine

learning is the size of the training data set. When this set is too
large, the algorithm will find patterns where they don’t exist, which
is called over-fitting. When the data set is not large enough, it will
be unable to find any patterns. Fine-tuning a predictive algorithm
to deliver reliable predictions is a craft that requires great insight
into the algorithm.

Unsupervised methods operate without labelled examples. These

algorithms trawl through the data to find patterns. Clustering
method detects groups of data in multi-dimensional data. This
technique is useful to segment customers using sets of charac-
teristics. Segmenting customers helps organisations to target their
Strategic Data Science 63

marketing. Unsupervised methods can also reduce the dimension-

ality of data. These techniques are helpful to discover otherwise
invisible aspects of the data. Factor analysis and structural equation
modelling are examples of methods to analyse social surveys and
find underlying patterns.
The output of these machine learning methods is very different
from traditional analysis. While scientists wear t-shirts with ele-
gant formulas, the outcomes of machine learning are often utterly
incomprehensible to humans. This aspect of machine learning can
be a problem as professionals hesitate to accept the outcomes of the
algorithm if they can’t see or understand the logic that the machine
uses to achieve its results. This problem highlights the need for data
science education among all professionals and not only specialists.
The third method, reinforcement learning is a technique that is used
in prescriptive analysis, which is the topic of the next section.

3.6 Prescribe Action

The last step in the continuum is where machines take over our
world, and human beings can relax and be served by their robot
slaves. This vision might seem science fiction, but automated
processes have been part of our lives for many decades.
Industrial systems need to be controlled by operators to ensure
it produces the outcomes we need. When, for example, a tank of
chemicals reaches a certain low level, a pump needs to be started
to fill it again, the operator needs to monitor the tank. Manual
systems require a human operator to review measurements of the
state of the system and take appropriate action. The first level of
automation uses a control loop to measure the level of the tank and
an automated valve.
Almost all contemporary manufacturing processes use these first-
generation control loops. The problem with these systems is that
Strategic Data Science 64

they rely on preset conditions that might not be suitable when ex-
ternal circumstances change. Traditional control systems measure
the present and take action when the measured value starts to trend
towards a set value. In effect, a control system consists of a network
of if-then statements to manage all the standard conditions. When
the environment deviates from the boundary conditions that the
system was designed to operate, the system will fail.

Machine learning can enhance traditional control systems by sim-

ulating the environment instead of relying on limited boundary
conditions. Reinforcement learning uses the principles of game
theory to find solutions that maximise outcomes. This method uses
a virtual carrot and stick as the algorithm finds a way to optimise
predefined conditions. Reinforcement learning is very close to
the way organisms learn. When an animal, human or otherwise,
behaves in a certain way is positively rewarded, the behaviour will
be reinforced and repeated in similar future situations.

This method simulates the environment it wants to control and

simulate millions of possible decisions to determine the most op-
timal choice. This method has been insanely successful in playing
complex board games such as go and chess. Reinforcement learning
is ideal for games because reality can be simulated perfectly.
Deploying these methods in real-world systems is much more
complicated because simulating a natural or social environment
is much more complicated than the ideal environment of a board
game.

3.7 Towards a Data-Driven

Organisation

The data science continuum provides a strategic map for organi-

sations that seek to become more data-driven. Each of the steps in
the continuum is equally important to the next level because these
Strategic Data Science 65

higher levels of complexity cannot be achieved without embracing

the lower levels. The most important aspect of the data science
continuum is that it summarises an evolutionary approach towards
becoming a data-driven organisation. As an organisation evolves
towards more complex forms of data science, the earlier stages
don’t become vestigial appendices but remain an integral part of the
data science strategy. All parts of this model are of equal relative
value.

Being data-driven is, however, more than a process of increas-

ing complexity. Evidence-based management requires the people
within the organisation to be data literate and work together to-
wards a common goal. The systematic aspect of data science needs
a formalised process to ensure sound outcomes. The increased
complexity of the analytical methods also requires an investment
in better tools and data infrastructure. The last chapter of this book
reflects on the non-technical aspects of data science and discusses
the people, process and systems required to become a data-driven
organisation. The book closes with a reflection on the limitations
of data science from a mathematical and ethical perspective.
4. The Data-Driven
Organisation
The previous two chapters describe some of the technical aspects
to consider when implementing data science in an organisation.
Merely focusing on the technicalities of analysing data is, however,
not enough to create value for an organisation. A data science
manager needs to manage people, systems and process to develop
a data-driven organisation.

Decision makers sometimes ignore even the most useful and aes-
thetic visualisations, even when the analysis is sound. Best practice
data science, as described in chapter two, is only the starting
point for creating a value-driven organisation. A critical aspect of
ensuring that managers use the results is to foster a data-driven
culture, which requires managing people.

To enable data science to flourish, the organisation needs to have a

well-established suite of IT systems to store and analyse data and to
present the results. A wide range of data science tools is available,
each playing a different role in the analytics value chain.

Every data science project starts with a problem definition that is

translated into data and code to define a solution. This problem is
injected into the data vortex until a solution is found. The process
of data science discusses the workflow of creating data products.

This final chapter discusses these three aspects of becoming a

data-driven organisation. To strategically implement data science
requires aligning people, systems and processes to optimise the
value that can be extracted from the available information. The
The Data-Driven Organisation 67

next three sections discuss these aspects of data science. This book
closes with some deliberations about the theoretical and ethical
limitations of data science.

4.1 People

When discussing the people related to a data-driven organisation,

we should not only mention the specialists that create the data
products. The members of the data science team possess the compe-
tencies shown in the Conway diagram in section 1.3. They also need
to be able to communicate the results of their work to colleagues or
clients and convince them to implement the findings.

Data science does not only exclusively happen within the spe-
cialised team. Each data project has an internal or external cus-
tomer that has a problem in need of an answer. The data science
team and the users of their products work together to improve the
organisation. This implies that a data scientist needs to understand
the basic principles of organisational behaviour and change man-
agement and be a good communicator. In reverse, the recipients
of data science need to have sufficient data literacy to understand
how to interpret and use the results.

4.1.1 The Data Science Team

Contemporary folklore portrays the data scientist as the slightly

socially awkward genius. The geek in the movies converts energy
drinks and junk food into computer code and solves a complex
problem with merely a few keystrokes. While much of data science
relies on the talents of highly skilled individuals, a strategic ap-
proach to creating value from data is more teamwork than it is the
domain of the lone genius. The ideal data science team combines
all the skills and expertise needed to do its task.
The Data-Driven Organisation 68

The Conway diagram, visualises the technical skills of the data

science team. To be able to embed the results into the organisation,
data scientists need people skills in addition to technical competen-
cies. These soft skills are not part of the Conway diagram because
they are common to all professions. Their absence from the picture
is no indication of their lesser importance. Section 4.1.3 discusses
the cultural aspects of being data-driven.

From the fact that a data scientist requires well-developed com-

petencies in writing code, it does not follow that the data science
function naturally forms part of the Information Technology (IT)
team. Writing code is traditionally seen as a competency of IT
professionals. Anyone outside this group that wants to maximise
the capabilities of the hardware and software they are provided is
often labelled as ‘shadow IT’.

Consultancy firm Gartner regularly releases predictions about how

they perceive the future of computing. A few years ago, they
suggested that a significant part of IT expenditure will occur
outside the traditional lines of responsibility.¹ As most functions
become ever more reliant on digital technology, traditional IT
skills are distributed around the organisation. The idea that only
IT professional write code is indeed incorrect from a data science
perspective because developing a spreadsheet is not much different
from developing software.

Centralising all IT management to one team is like having the

Human Resource (HR) manager conduct all performance reviews
or be responsible for hiring all people. Just like HR, information
technology skills need to be decentralised to maximise the value
we can extract from existing infrastructure.

The data science industry is frantically looking for so-called uni-

corns, which are people that have all three skills in the Conway
¹Gartner. (2011). Gartner Reveals Top Predictions for IT Organizations and Users
for 2012 and Beyond (Press Release).
The Data-Driven Organisation 69

diagram. The most effective way to achieve this ideal is to teach

existing subject-matter experts to write code. This approach creates
data scientists with domain knowledge and improves data literacy.
Embracing the shadow IT ethos and distributing coding skills
around the organisation is one part of developing a data-driven
organisation.
Data science can be positioned in two ways within an organisation.
A team can be embedded within a product or service group, or it
can be a dedicated team that services all parts of the organisation.
Smaller organisations will most likely have one specialised team
that services the organisation. A disadvantage of a centralised data
science team is that it creates a distance between the operational
managers and the data specialists, which complicates implementing
the results. Larger organisations have a workload that is sufficient
to justify more than one data science knowledge centre. Data
science competencies embedded with an operational team places
them within the action, which allows them to respond rapidly
to operational needs. This disadvantage of this method is that
expertise is fragmented across the organisation.²

4.1.2 Data Science Consumers

The main hurdle for creating a data-driven organisation is that not

everybody fully understands the results of sophisticated analysis.
Data scientists are responsible for producing aesthetic reports and
visualisations to ensure that the user grasps the end-product. When
professionals fail to understand the analysis, the results will be
ignored, and they revert to their intuitive understanding of the
problem.
Statistics can often be counter-intuitive, and some people have
a low level of trust in the mathematical wizardry of the data
scientist. Mark Twain popularised the phrase “Lies, damned lies,
²Caffo, Peng, & Leek (2018)
The Data-Driven Organisation 70

and statistics”. The complexity of statistics makes it an ideal vehicle

to use as a propaganda tool, which these days we call fake news.
Lying with statistics is easy when the receiver has a low level of
data literacy.³

Unfortunately, the physical and social world are inherently stochas-

tic in nature. In addition to this uncertainty, all measurements
introduce a level of random error. Statistics are thus the most
effective way to understand reality. We can never understand the
world with arbitrary mathematical precision and can only rely on
a statistical description. The most perfect knowledge of the world
we can ever have is statistical. When statistics are appropriately
applied, Mark Twain’s phrase should be reversed: “truth, absolute
truth and statistics”.

This fact places requirements on the consumers of the results of

data science. Data literacy is an inherently mathematical skill,
but unfortunately, many people have grown up fearing the most
beautiful of all sciences. This does not imply that all employees
need to do a mathematics course. The data scientist needs to create
aesthetic products that require only a common sense understanding
of the world. Unfortunately, this is only possible for the simplest of
business problems.

Moving towards artificial intelligence is a fool’s errand if we don’t

also invest in natural intelligence. Implementing a data science
strategy requires an investment in professional development to
increase data literacy. For complex data science to deliver value, the
whole of the organisation needs to be data literate at the appropriate
level.

Data literacy is the ability to identify, locate, interpret and evaluate

quantitative and qualitative information and communicate the
results. The scope of data literacy envelopes the whole of the
³Huff, D. (1985). How to Lie with Statistics (Reprint). Harmondsworth: Penguin
Books.
The Data-Driven Organisation 71

data science workflow described in section 4.3. The data literacy

curriculum consists of several areas of competency.

Employees need to be able to source information and understand

how data is governed. Employees undertaking analysis need to be
versed in research skills and know how to manipulate data. The
third aspect of data literacy are skills to interpret statistics. Visual-
ising information and presenting evidence to decision makers are
the last two data competencies.⁴

The extent to which employees need to have data competencies

depends on their position within the organisation. We expect a
much higher level of data literacy from technical professionals than
we would from administrative staff.

The reliance on electronic tools to interrogate data implies that

data literacy goes hand-in-hand with digital literacy. Users need
to be computer literate and able to use the various tools of the
trade confidently. Contemporary business intelligence tools have
simplified the level of skill required to extract and model multiple
sets of data. Having digital skills in isolation of data skills can, how-
ever, lead to problems because users draw incorrect conclusions
due to wrong methodologies or lack of validation, as discussed in
Conway’s danger zone.

Data science consumers are not only the colleagues of the data
scientist or the clients of a consultant, but they are also the con-
sumers of the products and services of the organisation. Businesses
and government agencies love to publish statistics to show their
customers or the community how awesome they are. Communi-
cating accurate statistics is an effective way to foster trust between
the organisation and its stakeholders. This method only works,
however, when the targeted people have sufficient data literacy to
understand the information. Most data communication is produced
⁴Australian Public Service Commission. (2018). Data literacy skills. Retrieved 2
March 2019, from apsc.gov.au.
The Data-Driven Organisation 72

by highly intelligent professionals who find it difficult to under-

stand the limitations of their less educated fellow citizens.
The European Union has published a framework or digital com-
petency of its citizens. This document recognises the importance
of digital literacy to maximise participation in society. This doc-
ument discusses information literacy, the ability to communicate
and collaborate, creating digital content, digital safety and solving
problems, as its five areas of competency.⁵
Closing the data literacy gaps between data science and its con-
sumers requires movement in two directions. Firstly, the analyst
needs to be a great communicator and convey their message in
convincing and straightforward language, as discussed in chapter
two. Secondly, the consumers of data science need to be educated
to a sufficient level to help them navigate the complex world of
statistics.

4.1.3 Data Science Culture

The data revolution has made some organisations realise that they
are data rich but information poor. Managers in these organisations
realise that they hold large amounts of data that is only used once.
The popularity of data science is to a large extent motivated by a
desire to use this dark data and become more data-driven.
The most complicated aspect of implementing a data science strat-
egy is to integrate the results of our analysis with every-day
business activities. A data-driven organisation is one where using
information to solve problems forms part of its culture. On the
surface, this can be achieved by updating the existing operating
procedure to include analysing data, but there is also a strong
human component to this transition.
⁵Carretero, S., Vuorikari, R., Punie, Y., European Commission, & Joint Research
Centre. (2017). DigComp 2.1 The Digital Competence Framework for Citizens with
Eight Proficiency Levels and Examples of Use.
The Data-Driven Organisation 73

The hundreds of published definitions of culture illustrate the

complexity of the concept of organisational culture. The main
reason for this confusion is that culture is often defined in abstract
terms for which there is no empirical way to study them. The most
pragmatic and productive way to discuss culture is to examine its
phenomena instead of trying to define its essence. A phenomeno-
logical approach investigates the visible aspects of a culture to learn
something about the invisible dimensions. The image in figure 14
uses the well-trodden iceberg analogy to list some of the visible and
invisible phenomena of organisational culture. This model helps to
explain what a data science culture means in practice and how to
recognise or create it.⁶

Figure 14: Phenomena of Organisational Culture.

Before we can say anything sensible about the data culture of an

organisation, we need to understand the artefacts. An artefact is
a visible expression of culture that can be observed, studied or
measured. A focus on the tangible aspects of culture also makes it
easier to change it because we can modify things we can measure.
⁶McShane, S., & Travaglione, T. (2005). Organisational Behaviour on the Pacific
Rim (Enhanced ed.). North Ryde N.S.W.: McGraw-Hill.
The Data-Driven Organisation 74

The physical structures of an organisation are the most concrete

artefacts. The office is a physical structure that expresses the culture
of an organisation. Is the office a cubicle maze or does it provide
an open and collaborative space to work in teams? The office is
also a place to distribute communication through works of art
and posters. A data-driven organisation displays visualisations and
other analytical results on walls or the intranet to inform employees
of exciting information.
An organisation is like a tribe that has its own specific rituals and
ceremonies. The word ritual is often reserved for a meaningless and
repetitive activity. Rituals and ceremonies are not merely useless
activities as they form an essential part of human social life. Also
outside the realm of religion, rituals are an indispensable tool to add
structure to social life. The induction of new staff is an initiation
ceremony that transforms the job applicant to an employee. This
ritual provides a strong signal to the novice about the culture of
their new place of work. A data-driven organisation introduces the
new employee to their relevant data structures and processes to
obtain and analyse this information.
The stories and legends within an organisation are the unwritten
narratives that circle around the office during informal meetings.
These stories are passed on and modified by the generations of
employees. If an organisation is data-driven, then at least some of
these stories would include stories about data and how it was used
to do something valuable. In a data-driven organisation, employees
not only communicate with data through formal reports, but they
also refer to it in informal settings.
The last artefact of culture is the language of an organisation. The
language of employees as expressed in reports, memos, emails and
so on, is a robust indicator of the extent to which data informs
decisions. Business reports, investment proposals or any other
formal document should use data as its foundation.
To assess the extent to which an organisation is data driven, you
The Data-Driven Organisation 75

should become a part-time anthropologist who uses participant-

observation to evaluate the four artefacts of culture. The more
these artefacts include reference to data, the more data-driven an
organisation is.
The beliefs, values and assumptions of organisational culture are
much harder to measure or observe than visible artefacts. Psycho-
metric surveys can provide some insight, but in corporate settings
these, more often than not, measure politically correct opinions,
instead of the actual attitudes of employees. Employees often
respond to what they believe should be the case, instead of what
is the case.
The invisible aspect of organisational culture and the visible arte-
facts have a two-way causal relationship. The intangible elements
of a culture influence how we behave and communicate and the
artefacts are thus a result of culture. This mechanism also works in
the reverse direction in that the artefacts modify our beliefs, values
and attitudes.
The collective minds of employees around the world shudder when
they hear managers utter the words ‘culture change’. Most cultural
change programs fail because managers don’t themselves display
the values they aspire to. Changing the culture to some ideal that a
manager reads about in an inspiring article is doomed to fail at the
best of times.
Changing the values and beliefs of employees in an organisation
is extremely hard and often leads to resistance. Sending employees
to development sessions to develop artificial values is in most case
not very productive. The easiest way to change an organisational
culture is to improve the visible signs. The intangible aspects of
culture will slowly but surely modify themselves to the messages
embedded within the artefacts. The physical structures change
the way employees interact with each other. As more new staff
are recruited, the percentage of employees that undergo the new
induction becomes greater. The more the formal and informal
The Data-Driven Organisation 76

language of the organisation includes references to data, the more

it will become part of its culture.

A data-driven manager can change an organisation by modifying

the artefacts of organisational culture and infuse them with reports
and visualisations. The more the artefacts include reference to data
and analysis, the more the beliefs, values and assumptions will
become driven by data.

4.2 Systems

Just like any other profession, a data scientist needs a suitable set of
tools to create value from data. A plethora of data science solutions
is available on the market, many of which are open source software.
Specialised tools exist for each aspect of the data science workflow.

There is no need to discuss the multitude of packages that are

available. Many excellent websites review the various offerings.
This section provides some thoughts on the use of spreadsheets
versus writing code and business intelligence platforms.

Spreadsheets are a versatile tool to analyse data that has prolifer-

ated into almost every aspect of business. This universal tool is,
however, not very suitable to undertake complex data science. One
of the perceived advantages of spreadsheets is that they contain
the data, the code and the output in one convenient file. This
convenience comes at a price as it reduces the soundness of the
analysis. Anyone who ever had the displeasure of reverse-engineer-
ing a spreadsheet will understand the limitations of spreadsheets.
In spreadsheets, it is not immediately clear which cell is the result of
another cell and which ones are original data. Many organisations
use spreadsheets as a single source of truth for corporate data,
which should be avoided if the information needs to be shared. Best
practice data science is to separate data, code and output.
The Data-Driven Organisation 77

As mentioned previously, the best way to breed data science

unicorns is to teach subject-matter experts to write analytical code.
Writing code, such as R or Python, is like writing an instruction
manual on how to analyse data. Anyone who understands the
language will be able to know how you derived your conclusions.
Modern data science languages can generate print-quality visu-
alisations and can output results in many formats, including a
spreadsheet or a stand-alone application.

The gold standard for data science coding is literate programming.

This technique combines the code with prose to enable the al-
gorithm to be fully understood. All computing languages include
the ability to add comments. Literate programming is a further
evolution of this technique. The results of literate programming are
most commonly a report or a deck of presentation slides.

Each language has its own methods to combine text with code.
RMarkdown, Jupyter Notebooks and Org Mode are popular sys-
tems to undertake literate programming. Once the code is written,
at the push of a button the computer generates a new report with
updated statistics and graphics. You can either choose to include or
exclude the code from the final result, depending on the expertise
of the reader.

Lastly, business intelligence tools are useful to disseminate the

results of a data science project but are not very useful to undertake
detailed analysis. A platform such as Power BI is a great system
to visualise the result of analysis because it provides very flexible
ways to slice-and-dice the data and visualise results. The analytical
capabilities of the platform are not very high but can be amended
by inserting code in Python or R to complement its capabilities.
The Data-Driven Organisation 78

4.3 Process

Chapter two mentioned the requirement for governance in data

science to ensure the outcomes of projects are sound. The process
of creating value from data follows an iterative workflow that
works from raw data to a finished project.⁷ The workflow starts
with defining a problem that needs solving (Figure 15). The next
step involves loading and transforming the data into a format that
is suitable for the required analysis. The data science workflow
contains a loop that consists of exploration, modelling and reflec-
tion that is repeated until the problem is solved or is shown to be
unsolvable.

Figure 15: Data science workflow.

The workflow for a data project is independent of the aspect of the

data science continuum under consideration. The same principles
apply to every type of analysis. For larger projects, formal project
management methods are advisable to control time, budget and
quality.

The sections below describe some salient aspects of each of the

⁷Wickham, H., & Grolemund, G. (2016). R for data science: Import, Tidy,
Transform, Visualize, and Model Data Sebastopol, CA: O’Reilly. Available at
r4ds.had.co.nz.
The Data-Driven Organisation 79

three phases in the data science workflow and present a case study
about reporting water quality to a board of directors.⁸

4.3.1 Define

The first step of a data science project is to define the problem. This
first step describes the problem under consideration and the desired
future state. The problem definition should not make specific
reference to available data or possible methods but be limited to
the issue at hand. An organisation could seek to optimise produc-
tion facilities, reduce energy consumption, monitor effectiveness,
understand customers and so on. A concise problem definition is
necessary to ensure that a project does not deviate from its original
purpose or is cancelled when it becomes apparent that the problem
cannot be solved.

The outcome of a data science project can be either actionable

intelligence or improved understanding of a business process, as
discussed in section 2.2.

The problem definition opens with a description of the current

situation and clearly identifies which aspect needs improvement or
more profound understanding. The problem statement concludes
with a summary of the ideal situation. For example:

A water utility reports water quality results to the board

of directors every month. The report contains a lot of
tables with results of laboratory testing. The members
of the board are not technical experts and don’t fully
understand what they mean. Another section of the
report shows water quality complaints, without indicat-
ing how they relate to water quality results. The board
⁸Prevos, P. (2015). Visualising water quality: A graphical index for drinking
water system performance. In OzWater. Adelaide: Australian Water Association.
The Data-Driven Organisation 80

has asked for a visualisation to report water quality

performance that combines the available data in one
succinct overview.

The definition of the project concludes with a description of a

possible method to solve the problem and the data sources needed
to implement the solution. Assuming all the required sources of
data are available, the analytical process can commence.

For the case study, the analyst decided to use a performance index
for each of the different aspects of the water supply system. A water
quality index is a dimensionless number that reflects the level of
performance compared to an ideal situation.

4.3.2 Prepare

The available data needs to be loaded and wrangled into the

required format before any analysis can take place. Influential data
scientist Hadley Wickham refers to this process as tidying data.⁹
Anecdotally, this phase of the project could consume up to eighty
per cent of the work effort, depending on the difference between
the available data and the required data.

Best practice in data science is to record every data set that is

considered for the project. A data book records every field extracted
from a source to ensure the context in which the data was created
is understood.

⁹Wickham and Grolemund (2016).

The Data-Driven Organisation 81

Table 1: Data science case study: available data.

Source Source Availability

Catchment Paper records Intermittent
Treatment plant SCADA Continuous
Network Laboratory Weekly
Customer experience Complaints Daily
Regulation Issues register Intermittent

For the water utility case study, several data sources are available
as shown in the table. The data scientist needs to decide which of
these sources solves the problem. In this example, the data from
the catchments is only available on paper, which makes it difficult
to analyse it algorithmically. A separate project would be required
to convert this source to electronic data. Other data sources are
available electronically and can be used for the project.

4.3.3 Understand

Once the data is available in a tidy format, the process of under-

standing the data can commence. The analytical phase consists
of a three-stage loop, the data vortex, that is repeated until the
required results are achieved or evidence becomes available that
the objectives cannot be met. These three stages are explore, model
and reflect.

4.3.3.1 Explore

The best method to start analysing data is to explore the data

to understand its relationship to reality. Generating descriptive
statistics such as averages, ranges, correlations and so on provides
a quick insight into the data. Relying on numerical analysis alone
can, however, deceive because very different sets of data can result
in the same values.
The Data-Driven Organisation 82

Justin Matejka and George Fitzmaurice from AutoDesk demon-

strated how very different sets of data can have almost the same
summary statistics (Figure 16). Each of these six visualisations
shows that these sets of data have very different patterns. When,
however, analysing this data without visualising it, the mean values
of x and y, and their correlations, are almost exactly the same for
all six subsets. In their paper, they presented an algorithm that
generates every possible pattern with the same summary values,
six of which are shown in the illustration.¹⁰

Figure 16: Six patterns with very similar summary statistics.

Another reason to use visualisations to explore data is to reveal

anomalies, such as unnatural spikes or outliers. A sudden increase
and decrease in physical measurements are often caused by issues
with measurement or data transmission instead of actual changes
in reality. These spikes need to be removed to ensure the analysis is
¹⁰Matejka, J., & Fitzmaurice, G. (2017). Same Stats, Different Graphs: Generating
Datasets with Varied Appearance and Identical Statistics through Simulated An-
nealing. In Proceedings of the 2017 CHI Conference on Human Factors in Computing
Systems - CHI ’17 (pp. 1290–1294). Denver, Colorado, USA: ACM Press. DOI
10.1145/3025453.3025912.
The Data-Driven Organisation 83

reliable. Anomalies in social data such as surveys could be subjects

that provide the same question to all answers, discussed in the
previous chapter.
Detecting outliers and anomalies and removing these from a data
set increases the reliability of the analysis. Not all anomalies in a set
of data are necessarily suspicious, and care should be taken before
removing data. The reasons for removing any anomalous data
should be documented so that the analyses remains reproducible.
The exploratory analysis in the case study involved generating
time series charts for all relevant data to understand the statistical
distributions. What are the maximum and minimum values? How
variable are these observations, and other basic statistics provide
insight into the shape of the used data.

4.3.3.2 Model

After the analyst has a good grasp of the variables under con-
sideration, the actual analysis can commence. Modelling involves
transforming the problem statement into mathematics and code, as
described in the previous chapter.
Every model of the world is bounded by the assumptions contained
within it. Statistician George Box is famous for stating that “all
models of reality are wrong, but some are useful”. Since data science
is not a science in the sense that we are seeking the truth, a useful
model is all we need.
When modelling the data, the original research question always
needs to be kept in mind. Exploring and analysing data with-
out a specific purpose can quickly lead to wrong conclusions.
Just because two variables correlate does not imply that there is
a logical relationship. A clearly defined problem statement and
method prevent data dredging. The availability of data and the
ease of extracting this information makes it easy for anyone to find
relationships between different sources of information.
The Data-Driven Organisation 84

A good general rule when analysing data is to distrust your method

when you are able to confirm your hypothesis easily. If this is the
case, using triangulation through a different method helps to verify
the results.
The modelling in this case study involved developing a demerit
point system. Perfect water quality is almost always defined by
lower and upper limits for a particular parameter. Perfect water
has no colour, has a minimum level of chlorine, and so on. This
information forms the basis of a points system. When, for example,
water has a measurable colour, the index is reduced by a certain
amount of points. A decision rule was designed for each parameter
in the model.

4.3.3.3 Reflect

Before the results of an analysis can be communicated, domain

experts need to review the outcomes to ensure they make sense
and indeed solve the problem stated in the definition. The reflection
phase should also include the customer to ensure that the problem
statement is being solved to their satisfaction.
Visualisation is a quick method to establish whether the outcomes
make sense by revealing apparent patterns. Another powerful tech-
nique to reflect on the results is sensitivity analysis. This technique
involves changing some of the assumptions to test the model
responds as expected. Mathematical models are often complex
where the relationship between variables is not clearly understood.
Sensitivity analysis helps to understand these relationships by using
extreme for certain variables and then observe the effect on the
model.
Concerning the case study, the modelling required a lot of reflec-
tion on the chosen model and required many iterations of the data
vortex. Developing a performance index is a balancing act to ensure
each parameter is appropriately weighted. The purpose of the index
The Data-Driven Organisation 85

was to provide the board with salient information so they can ask
targeted questions during meetings. The reflection phase always
needs to reflect on the purpose and ensure it is achieved.

4.3.4 Communicate

The last, and arguably the hardest phase of a data science project
is to communicate the results to the users. In most cases, the users
of the data product are not specialists with a deep understanding
of data and mathematics. The difference in skills between the data
scientist and the user of their products requires careful communi-
cation of the results.

Water System Index visualisation

Detailed reports and visualisations need to not only provide an

accurate description of the outcomes, but they also need to be able
to convince the reader. The most critical aspect of successfully
communicating the solution to the problem is to ensure that the
consumers of the results understand them and are willing to use
them to solve the problem. As much as data science is a systematic
The Data-Driven Organisation 86

process, it is also a cultural process that involves managing the

internal change in the organisation, discussed in section 4.3.1.

The final result of this case study is a spatial visualisation of all

the water treatment plants within the service region of the water
utility. Each of the circles on the chart visualises the result of the
index using a diverging scheme from red to green. The size of the
circle is proportionate to the logarithm of the number of customers
in each town to emphasise the greater importance of larger systems.
To enable reproducibility, clicking on the circles provides the user
with all the internal workings and original data so that the root
cause of non-performance can be identified.

4.4 The Limitations of Data Science

The problems that data science can solve seem to be limitless

when reading some of the articles on the web and in literature.
There are, however, theoretical and practical limitations to what a
computational analysis can achieve. Besides limits to what we are
capable of doing, there are also ethical limits to what we should be
doing. This final section reflects on some of the boundaries of data
science.

4.4.1 The Limits of Computation

The limitations of computation relate to some complex and deeply

philosophical issues concerning the limits of human knowledge.
This short introduction to this problem looks at the restrictions
of measuring physical and social processes and the limitations of
algorithms.
The Data-Driven Organisation 87

4.4.1.1 Limitations of Measurement

The first limitation relates to the fact that our collection of data will
always be an incomplete description of reality. In a physical system,
choose which points to measure, at which frequency, by which
method and so on. We need to make a lot of choices before we can
access data. Even more so with social data, all our measurements
are only indirect expressions of the reality we seek to explain.

As discussed in section 2.3, measurements are never a perfect copy

of the reality it describes. There is no such thing as raw data because
all measurement uses assumptions about the reality it represents.

These limitations of data restrict what we can know about the

future. Predictions all use the same principle to achieve a result.
Patterns in the data about the past are used to extrapolate the
current situation to a plausible future. As our data is, however,
always incomplete, these models will quickly sharply produce

Complex systems such as the weather will never be predicted over

the long term due to the inherent gaps in data from a spatial
perspective. Dynamic systems are highly dependent on initial
conditions. A slight measurement error will be amplified over time
and render long-term predictions impossible.

Social systems are even more complicated than dynamic physical

systems. Even though big data is quite successful at modelling
aspects of human behaviour, it is incapable of providing a broad
model of our actions. Amazon might be able to predict which brand
of laptop I am most likely to purchase, there is not much more they
can do with the information I volunteer to them.

Although it might in principle be possible to measure everything,

the simple cost of doing so is prohibitive. Certain aspects of reality
are intrinsically hard to measure. A seemingly simple parameter,
such as the presence of pathogens, takes a lot of effort to deter-
mine. Even if we succeed in measuring vast amounts of data, the
The Data-Driven Organisation 88

computational cost of processing this information needs to relate

to the anticipated benefits reasonably.

4.4.1.2 Limitations of Algorithms

Computational methods to solve problems are potent methods to

answer questions that we not long ago not even dared asking. There
are, however, specific problems that even the smartest algorithms
cannot solve.

An example of an intrinsically hard problem is the travelling

salesman. An organisation has salespeople that need to visit a large
number of cities. Salespeople seek to minimise the amount of time
they spend on the road and thus they need to know the shortest
possible routes between all towns. Although some algorithms exist
for this problem, they only provide estimates. To determine the
actual shortest path requires an astronomical amount of computing
time. If the problem contains a high enough number of cities, the
computational time necessary to find the solution could be longer
than the age of the universe.

4.4.1.3 Beyond the Limitations

Some of these limitations might be circumvented in the future,

but some of these issues are fundamental problems without a
solution. The problem of incomplete information will never be
fully resolved as the practical difficulties are insurmountable. The
power of computers has grown exponentially for the past fifty
years, and quantum computing will be the next leap in capabilities.
The problems with algorithmic complexity will, however, not be
resolved by even the fastest quantum computer imaginable.

The limitations listed in this section are profound and, as far as we

can gather from our current perspective, are unassailable. These
limitations are not stopping us from doing a lot more with data
The Data-Driven Organisation 89

than most organisations are capable of achieving at this moment.

Even within the constraints of solving problems with computa-
tional analysis, there are still many things we can accomplish with
algorithms.
The limitations of data science not only extend to what data science
could do but also to what it ought to do. Analysing data is an
inherently human activity which means it always has an ethical
dimension. The ethics of data science places limitations on what
info we collect, how we disseminate this information and what we
do with the results of an analysis.

4.4.2 Ethical Data Science

Technology gives humanity capabilities far beyond what we are

naturally capable of. Machines make us stronger, but they can
cause physical damage to people and the environment. Computers
make us smarter, but they can also cause social damage. With this
great power comes great responsibility. Donald Knuth is one of
the original computer gurus and an influential scholar in computer
science. Knuth questioned the limits of data science from an ethical
perspective when interviewed by the New York Times.¹¹

“I am worried that algorithms are getting too prominent

in the world … It started out that computer scientists
were worried nobody was listening to us. Now I’m
worried that too many people are listening.”

Knuth is not alone in his concerns about how algorithms can

influence our lives. Although algorithms themselves are always
ethically neutral, the actions that flow from them do have conse-
quences. We defined data science as a tool to influence reality in
¹¹Roberts, S. (2018). The Yoda of Silicon Valley. The New York Times. Retrieved
25 February 2019 from nytimes.com.
The Data-Driven Organisation 90

section 2.2. Algorithms can have a real impact on the world, which
implies that they are subjects to ethics.

This section provides some ethical guidelines that managers can use
in any discussion about whether a specific use of data is ethically
justified. Chapter two defined good data science as being useful,
sound and aesthetic. Perhaps we need to add a fourth aspect to this
trivium and insist that data science also needs to be ethical.

The power of contemporary approaches to analysing data can lead

to adverse consequences, which means that the results of data
science have moral implications. Data ethics is the last section of
this book because it creates the boundary condition within which
all activities should take place. This section discusses a framework
for the ethical use of data and how practitioners can apply moral
principles to their work.

The collection, storage and analysis of data that relates to humans

needs to be undertaken ethically. This statement is almost a tau-
tology, but the recent spate of instances where digital companies
betrayed the trust of their customers suggests that there is a need for
reflection on this issue. The privacy of digital consumers has been
violated numerous times and some algorithms potentially harm the
consumer. The ethics of data science goes beyond privacy as the
outcomes of algorithms can negatively impact our lives. Algorithms
can reinforce discrimination and reduce the power of buyers in a
free market.¹²

Ethics is a nebulous term, especially for data scientists who are

more comfortable interpreting objective numbers than dealing with
subjective moral reasoning. The ethics of data science is one of
those problems that data science itself cannot solve. There are
no algorithms that can help us in deciding whether something is
ethical. Ethical issues are best resolved through contemplation and
¹²O’Neil, C. (2016). Weapons of Math Destruction: How Big Data Increases
Inequality and Threatens Democracy. New York: Crown.
The Data-Driven Organisation 91

debate. Philosophers have put forward several competing models

for ethical reasoning, none of which has gained anything close to
universal acceptance.
The ethical perspective is closely related to the legal point of view,
but they are not necessarily the same. From the fact that an action is
legal you cannot conclude that it is also ethical, and vice versa. The
pace of technological change is much faster than legislative change,
which means that there is a need for ethical reflection on data
science. Given that the legal framework for data differs country by
country or even states within countries, this section only discusses
the principles of ethical data science.
The ethics of gathering, storing and interpreting information that
describes the lives of people has been well-developed in traditional
social research. Social scientists have developed these principles in
response to negative experiences with unethical research. Three
crucial ethical principles that apply to data science are:¹³

• Informed consent
• Avoiding harm in collecting data
• Doing justice to participants in analysing data

4.4.2.1 Informed Consent

In social research, all subjects must consent to their data being

collected and analysed. When we ask people to complete surveys,
subjects can choose which information they provide, or abstain
from providing information about themselves. The principle of
informed consent ensures that participants are neither deceived nor
coerced into surrendering information about their personal lives.
Informed consent applies as much to scholarly research as it does
to business operations. Recent high-profile cases with Facebook
¹³Bryman, A., & Bell, E. (2011). Ethics in Business Research. In Business Research
Methods (3rd ed.). Oxford: Oxford University Press.
The Data-Driven Organisation 92

contradict this notion and suggest that the community expects

that all governments, researchers and businesses seek consent
for any information they collect about people. Informed consent
relies on three conditions. Firstly, the subject of the information
gathering needs to be informed. Secondly, consent must be given
voluntarily without undue influence. Lastly, the person consenting
must have the capacity to consent, which requires them to be able
to understand the consequences of opting in. Many of the ethical
breaches published in the news relate to breaching either of these
three conditions of informed consent.

4.4.2.2 Avoiding Harm in Collecting Data

In social research, harm in collecting data includes any physical

or psychological consequences for the participant, such as their
development or self-esteem, inducing stress and introducing risk
to career prospects or future employment. Privacy is the most
significant concern regarding this principle. Most countries have
well-established privacy legislation that defines how organisations
deal with data.

One of the most common ways to protect privacy is to de-identify

data. Making data anonymous involves stripping it from person-
ally-identifying information such as names and addresses. Aggre-
gating data is one of the methods to remove personal data. By
aggregating data into groups, any link with individuals seems to
be removed. However, when these aggregated sets of data are
combined with other sets, the individual entries can in principle
be decomposed.

In an age of ever-increasing availability and size of data, anonymity

becomes almost impossible to guarantee. A tension exists between
protecting privacy and gaining societal benefits from data.¹⁴
¹⁴Oppermann, I. (2017). Data Sharing Frameworks (Technical White Paper).
Sydney: Australian Computer Society.
The Data-Driven Organisation 93

4.4.2.3 Doing Justice

Doing justice to participants implies that any analysis and inter-

pretation of information about customers reflects their interests.
Doing justice to the subjects of the analysis is algorithmic fairness.
This concept expresses the risk of potential harm to people or the
natural environment through the way the data is analysed.

Algorithmic fairness exists in two areas. Machine learning algo-

rithms use information from the past to predict possible futures.
Because of this foundation in the past, any machine learning
algorithm will amplify any bias that exists within the training data.
For example, a recruitment algorithm that searches for candidates
on LinkedIn for an engineering job is likely to be gender-biased
because of the past societal preference for male engineers.

The second aspect of algorithmic fairness relates to an imbalance

in power between the subject of data science and the organisation
that uses the algorithm. Online shops can use machine learning
to present every visitor to their website with a different price.
Algorithms can maximise the amount that customers pay for a
service. Although perfectly legal, this is only possible due to an
imbalance in power between the seller and the buyer.

Algorithmic fairness is a topic of much discussion in the industry

and media. The new capabilities of machine learning will require
society to rethink its moral and legal frameworks as we enter
previously unknown territory.

4.4.2.4 Don’t be Creepy

These three principles provide a starting point for starting a discus-

sion about ethical data science. Ethical principles, instead of values
and fixed rule are a useful tool to help clarify any potential ethical
issues in a data science project.
The Data-Driven Organisation 94

These three principles can be summarised into a simple principle:

“Don’t be creepy”. When you have a data science proposal and are
not sure about its ethical status then perhaps you should ask a few
friends about how they would feel about being a subject in this
analysis.
5. References
Anderson, C. (2015). Creating a Data-Driven Organization: Practi-
cal Advice from the Trenches. Sebastopol, CA: O’Reilly Media
Inc.

Bennett, C.M., Baird, A.A., Miller, M.B., & Wolford, G.L. (2009).
Neural correlates of interspecies perspective taking in the post-
mortem Atlantic Salmon: an argument for multiple compar-
isons correction. NeuroImage, 47, S125. DOI 10.1016/S1053-
8119(09)71202-9.

Bryman, A., & Bell, E. (2011). Ethics in Business Research. In Busi-

ness Research Methods (3rd ed.). Oxford: Oxford University
Press.

Burkov, A. (2019). The Hundred-Page Machine Learning Book.

LeanPub.

Caffo, B., Peng, R., & Leek, J.T. (2018). Executive Data Science.
A Guide to Training and Managing the Best Data Scientists.
LeanPub.

Carretero, S., Vuorikari, R., Punie, Y., European Commission, &

Joint Research Centre. (2017). DigComp 2.1 The Digital Com-
petence Framework for Citizens with Eight Proficiency Levels
and Examples of Use.

Clegg, B. (2017). Big Data: How the Information Revolution Is

Transforming Our Lives. Icon Books.

Davenport, T.H., & Patil, D.J. (2012). Data scientist: The sexiest job
of the 21st century. Harvard Business Review, 90(10), 70–76.
References 96

Davis, P.J., & Hersh, R. (1990). Descartes’ Dream. The World Ac-
cording to Mathematics. London: Penguin.

DeVellis, R.F. (2011). Scale Development: Theory and Applications

(3rd ed.). SAGE Publications.

Harrower, M., & Brewer, C.A. (2003). ColorBrewer.org: An Online

Tool for Selecting Colour Schemes for Maps. The Cartographic
Journal, 40(1), 27–37. DOI 10.1179/000870403235002042.

Huff, D. (1985). How to Lie with Statistics (Reprint). Harmondsworth:

Penguin Books.

Jaworski, B.J. (2011). On managerial relevance. Journal of Market-

ing, 75(4), 211–224. doi 10.1509/jmkg.75.4.211.

Jones, G. E. (2007). How to Lie with Charts (2nd ed). Santa Monica,
Calif: LaPuerta.

Kelleher, J.D., & Tierney, B. (2018). Data science. Cambridge, Mas-

sachusetts: The MIT Press.

Lankow, J., Ritchie, J., & Crooks, R. (2012). Infographics: The Power
of Visual Storytelling. Hoboken, N.J: John Wiley & Sons, Inc.

Lee, N., & Lings, I. (2008). Doing Business Research: A Guide to

Theory and Practice. London: SAGE.

Lindström, M. (2010). Buy.Ology: Truth and Lies About Why We

Buy. New York: Broadway Books.

Matejka, J., & Fitzmaurice, G. (2017). Same Stats, Different Graphs:

Generating Datasets with Varied Appearance and Identical
Statistics through Simulated Annealing. In Proceedings of the
2017 CHI Conference on Human Factors in Computing Systems
- CHI ’17 (pp. 1290–1294). Denver, Colorado, USA: ACM Press.
DOI 10.1145/3025453.3025912.
References 97

McShane, S., & Travaglione, T. (2005). Organisational Behaviour on

the Pacific Rim (Enhanced ed.). North Ryde N.S.W.: McGraw-
Hill.

O’Neil, C. (2016). Weapons of Math Destruction: How Big Data

Increases Inequality and Threatens Democracy. New York:
Crown.

Oppermann, I. (2017). Data Sharing Frameworks (Technical White

Paper). Sydney: Australian Computer Society.

Pletser, V. (2012). Does the Ishango Bone Indicate Knowledge of the

Base 12? An Interpretation of a Prehistoric Discovery, the First
Mathematical Tool of Humankind. Eprint ArXiv:1204.1019.

Prevos, P. (2017). Lifting the ‘Big Data’ Veil. Creating Value through
Applied Data Science. Water E-Journal, 2(1), 1–5. DOI 10.21139
/wej.2017.008.

Prevos, P. (2015). Visualising water quality: A graphical index for

drinking water system performance. In OzWater. Adelaide:
Australian Water Association.

Prevos, P. (2012). The incompleteness theorem of performance

measurement in service delivery. In World Business Capabil-
ity Congress. Auckland: Centre for Organisational Excellence
Research (COER).

Siegel, E. (2016). Predictive Analytics: The Power to Predict Who Will

Click, Buy, Lie, or Die. Wiley.

Taylor, F.W. (1997). The Principles of Scientific Management. Mine-

ola, N.Y: Dover Publications.

Tufte, E. R. (1997). Visual Explanations: Images and Quantities,

Evidence and Narrative. Cheshire, Conn.: Graphics Press.

Welch, J. (2001). Straight from the Gut. Warmer Books.