Strategy Analytics Summary
Provost, F., & Fawcett, T. (2013). Data science for business. O’Reilly.
Strategy Analytics course 325132-M-6
This document provides a overview of Strategy Analytics, bridging the gap between technical data science
algorithms and high-level business strategy. It serves as a guide for understanding how organizations can
leverage Data-Driven Decision-Making (DDD) to transition from intuition-based management to evidence-
based precision, ultimately securing a sustainable competitive advantage.
The summary begins with the foundations of data strategy (Chapters 1 & 2). It defines data science not
merely as a technical discipline but as a strategic asset, illustrated by the Capital One case study where data
acquisition was prioritized over short-term profit. These chapters introduce the CRISP-DM cycle, a
standard process for data mining and the concept of "Analytical Engineering," which is the art of
decomposing vague business problems into specific, solvable tasks like classification or regression.
The core technical methodologies are explored in Chapters 3 through 6. The text details Supervised
Segmentation, explaining how Decision Trees use Entropy and Information Gain to partition data into
homogeneous groups. It contrasts this with Parametric Modeling (Chapter 4), distinguishing between
Linear Regression, Logistic Regression, and Support Vector Machines (SVMs) based on their specific
objective functions. A critical focus is placed on Overfitting (Chapter 5), the danger of modeling random
noise rather than signal and the techniques used to avoid it, such as Cross-Validation. Chapter 6 expands
into Similarity, applying distance metrics to both Nearest Neighbor prediction and Unsupervised
Clustering.
Moving from creation to assessment, Chapters 7 and 8 focus on Model Evaluation. The text argues that
simple accuracy is often misleading in business contexts due to unbalanced classes and unequal costs.
Instead, it advocates for the Expected Value Framework and visual tools like ROC Curves and Profit
Curves to assess a model's true economic impact.
Advanced applications are covered in Chapters 9 through 12. These sections explore Generative Models
like Naive Bayes, techniques for Text Mining (transforming unstructured text into TFIDF vectors), and Co-
occurrence grouping for market basket analysis. Chapter 11 revisits Analytical Engineering to tackle
complex problems like Churn and Uplift Modeling, emphasizing the need to isolate causal influence from
simple correlation.
Finally, the document concludes with the managerial and ethical dimensions (Chapters 13 & 14). It
outlines how to manage data science teams, defining the ideal "T-shaped" data scientist who combines
deep technical skills with broad business acumen. The summary ends by addressing the "Virtuous Cycle"
of data that sustains competitive moats and the critical ethical responsibilities regarding privacy,
transparency, and algorithmic bias. This holistic view ensures that the reader understands not just how to
build a model, but how to deploy it responsibly to drive business value.
,1. CONTENTS
1. CHAPTER 1, “INTRODUCTION: DATA-ANALYTIC THINKING”, AND LECTURE 1..........................................3
2. CHAPTER 2, "BUSINESS PROBLEMS AND DATA SCIENCE SOLUTIONS," AND LECTURE 1..........................6
3. CHAPTER 3, "INTRODUCTION TO PREDICTIVE MODELING: FROM CORRELATION TO SUPERVISED
SEGMENTATION," AND LECTURE 2............................................................................................................9
4. CHAPTER 4, "FITTING A MODEL TO DATA," AND LECTURE 2.................................................................12
5. CHAPTER 5, "OVERFITTING AND AVOIDANCE," AND LECTURE 3...........................................................15
6. CHAPTER 6, "SIMILARITY, NEIGHBORS, AND CLUSTERS," AND LECTURE 3............................................18
7. CHAPTER 7 "DECISION ANALYTIC THINKING I: WHAT IS A GOOD MODEL?" AND LECTURE 4.................21
8. CHAPTER 8, “VISUALIZING MODEL PERFORMANCE" AND LECTURE 4...................................................25
9. CHAPTER 11, “DECISION ANALYTIC THINKING II: TOWARD ANALYTICAL ENGINEERING,” AND LECTURE 4
...............................................................................................................................................................28
10. CHAPTER 9, "EVIDENCE AND PROBABILITIES," AND LECTURE 5..........................................................31
11. CHAPTER 10, "REPRESENTING AND MINING TEXT," AND LECTURE 5..................................................34
12. CHAPTER 12, "OTHER DATA SCIENCE TASKS AND TECHNIQUES," AND LECTURE 5...............................37
13. CHAPTER 13, "DATA SCIENCE AND BUSINESS STRATEGY," AND LECTURE 6..........................................40
14. CHAPTER 14, "CONCLUSION," AND LECTURE 6...................................................................................43
,1. CHAPTER 1, “INTRODUCTION: DATA-ANALYTIC THINKING”, AND LECTURE 1
1. The Core Definitions: Data Science vs. Data-Driven Decision Making
To understand Strategy Analytics, you must distinguish between the activity of analysis and the strategic
approach to decision-making.
Data-Driven Decision-Making (DDD): This refers to the practice of basing decisions on the
analysis of data rather than purely on intuition or experience.
o The Value of DDD: Research by Brynjolfsson et al. shows that firms adopting DDD are
statistically more productive (by 4%–6%) and have higher returns on assets and equity than
firms that rely on intuition.
o Types of Decisions: DDD applies to two main types of decisions:
1. Discoveries: Analyzing data to find new patterns (e.g., Walmart discovering that
strawberry Pop-Tarts sell 7x more before a hurricane).
2. Repetitive Decisions: Improving the accuracy of massive scale, routine decisions
(e.g., MegaTelCo predicting customer churn for millions of accounts).
Data Science: This involves the principles, processes, and techniques for understanding phenomena
via the automated analysis of data. It is the extraction of knowledge.
o Data Science vs. Data Engineering: Data Science focuses on extracting knowledge (the
"science"). Data Engineering focuses on the hardware, software, and pipelines to process
massive amounts of data (the "plumbing," like Hadoop or MongoDB). While Big Data
technologies (Volume, Variety, Velocity) support data science, using them does not
automatically mean you are doing data science.
2. Data as a Strategic Asset: The Capital One Case
A central theme is viewing data not just as a byproduct of business, but as a strategic asset that requires
investment. This is best illustrated by the Signet Bank (Capital One) case study.
The Problem: In the 1990s, banks offered credit cards with uniform pricing because they lacked
data to differentiate customers. They could not identify which customers were profitable and which
were high-risk.
The Strategy: Signet Bank realized that if they could model profitability, they could offer better
terms to good customers ("skim the cream") and avoid bad ones. However, they lacked the data to
build these models because they had never offered varied terms before.
The Solution (Data as Asset): They treated data as an asset to be acquired. They deliberately offered
credit with random terms to random customers. This resulted in immediate financial losses (bad
loans), but these losses were viewed as the cost of data acquisition.
The Result: The data generated allowed them to build superior predictive models for profitability,
leading to the spin-off of Capital One, which became a market leader by tailoring products to specific
customer risk profiles.
Exam Takeaway: You may need to sacrifice short-term profit to generate the data necessary to build a
competitive advantage.
, 3. Fundamental Data Mining Tasks
You must be able to map a business problem to one of the specific data mining tasks.
Classification: Predicting which of a small set of classes an individual belongs to (e.g., "Will this
customer churn?" -> Yes/No).
Regression (Value Estimation): Predicting a numerical value for an individual (e.g., "How much
will this customer use the service?").
Similarity Matching: Identifying similar individuals based on data (e.g., IBM finding companies
similar to their best customers to generate leads).
Clustering: Grouping individuals by similarity without a specific purpose or target variable (e.g.,
"Do our customers form natural groups?").
Co-occurrence Grouping: Finding associations between entities based on transactions (e.g.,
"People who bought X also bought Y").
Profiling: Characterizing typical behavior (e.g., "What is the normal credit card usage for this
segment?"). This is often used for anomaly/fraud detection.
Link Prediction: Predicting connections between data items (e.g., "Since you and Karen have 10
mutual friends, you should be friends").
Causal Modeling: Helping understand what events or actions actually influence others (e.g., "Did
the ad cause the purchase, or would they have bought it anyway?").
4. Supervised vs. Unsupervised Methods
This is a critical technical distinction in Strategy Analytics.
Supervised Learning:
o Definition: There is a specific target variable (outcome) you are trying to predict.
o Requirement: You need labeled historical data where the value of the target is known.
o Examples: Classification (Target = Categorical, e.g., Churn/No Churn) and Regression
(Target = Numerical, e.g., Revenue).
o Evaluation: Can be mathematically evaluated because we can compare predictions to actual
known outcomes.
Unsupervised Learning:
o Definition: There is no target variable. The goal is exploration or pattern finding.
o Examples: Clustering, Profiling, Co-occurrence grouping.
o Evaluation: Harder to evaluate because there is no "correct" answer to compare against.
Exam Tip: If a question asks "Can we find groups of customers who are likely to cancel?", this is
Supervised (Target = Cancel). If it asks "Do our customers fall into natural groups?", this is Unsupervised
(No target).
5. The Data Mining Process (CRISP-DM)
Data science is not a linear software development cycle; it is an exploratory cycle codified by the CRISP-
DM framework.
1. Business Understanding: Defining the problem to be solved. Creativity is essential here to recast
business problems as data science problems.
2. Data Understanding: Estimating the costs and benefits of data sources. Determining if the data
matches the problem (e.g., checking for biases).
3. Data Preparation: Converting data into a tabular format, removing missing values, and preventing
"leaks" (variables that give away the target but won't be available in production).
4. Modeling: Applying data mining techniques (algorithms) to the data to extract patterns.
Provost, F., & Fawcett, T. (2013). Data science for business. O’Reilly.
Strategy Analytics course 325132-M-6
This document provides a overview of Strategy Analytics, bridging the gap between technical data science
algorithms and high-level business strategy. It serves as a guide for understanding how organizations can
leverage Data-Driven Decision-Making (DDD) to transition from intuition-based management to evidence-
based precision, ultimately securing a sustainable competitive advantage.
The summary begins with the foundations of data strategy (Chapters 1 & 2). It defines data science not
merely as a technical discipline but as a strategic asset, illustrated by the Capital One case study where data
acquisition was prioritized over short-term profit. These chapters introduce the CRISP-DM cycle, a
standard process for data mining and the concept of "Analytical Engineering," which is the art of
decomposing vague business problems into specific, solvable tasks like classification or regression.
The core technical methodologies are explored in Chapters 3 through 6. The text details Supervised
Segmentation, explaining how Decision Trees use Entropy and Information Gain to partition data into
homogeneous groups. It contrasts this with Parametric Modeling (Chapter 4), distinguishing between
Linear Regression, Logistic Regression, and Support Vector Machines (SVMs) based on their specific
objective functions. A critical focus is placed on Overfitting (Chapter 5), the danger of modeling random
noise rather than signal and the techniques used to avoid it, such as Cross-Validation. Chapter 6 expands
into Similarity, applying distance metrics to both Nearest Neighbor prediction and Unsupervised
Clustering.
Moving from creation to assessment, Chapters 7 and 8 focus on Model Evaluation. The text argues that
simple accuracy is often misleading in business contexts due to unbalanced classes and unequal costs.
Instead, it advocates for the Expected Value Framework and visual tools like ROC Curves and Profit
Curves to assess a model's true economic impact.
Advanced applications are covered in Chapters 9 through 12. These sections explore Generative Models
like Naive Bayes, techniques for Text Mining (transforming unstructured text into TFIDF vectors), and Co-
occurrence grouping for market basket analysis. Chapter 11 revisits Analytical Engineering to tackle
complex problems like Churn and Uplift Modeling, emphasizing the need to isolate causal influence from
simple correlation.
Finally, the document concludes with the managerial and ethical dimensions (Chapters 13 & 14). It
outlines how to manage data science teams, defining the ideal "T-shaped" data scientist who combines
deep technical skills with broad business acumen. The summary ends by addressing the "Virtuous Cycle"
of data that sustains competitive moats and the critical ethical responsibilities regarding privacy,
transparency, and algorithmic bias. This holistic view ensures that the reader understands not just how to
build a model, but how to deploy it responsibly to drive business value.
,1. CONTENTS
1. CHAPTER 1, “INTRODUCTION: DATA-ANALYTIC THINKING”, AND LECTURE 1..........................................3
2. CHAPTER 2, "BUSINESS PROBLEMS AND DATA SCIENCE SOLUTIONS," AND LECTURE 1..........................6
3. CHAPTER 3, "INTRODUCTION TO PREDICTIVE MODELING: FROM CORRELATION TO SUPERVISED
SEGMENTATION," AND LECTURE 2............................................................................................................9
4. CHAPTER 4, "FITTING A MODEL TO DATA," AND LECTURE 2.................................................................12
5. CHAPTER 5, "OVERFITTING AND AVOIDANCE," AND LECTURE 3...........................................................15
6. CHAPTER 6, "SIMILARITY, NEIGHBORS, AND CLUSTERS," AND LECTURE 3............................................18
7. CHAPTER 7 "DECISION ANALYTIC THINKING I: WHAT IS A GOOD MODEL?" AND LECTURE 4.................21
8. CHAPTER 8, “VISUALIZING MODEL PERFORMANCE" AND LECTURE 4...................................................25
9. CHAPTER 11, “DECISION ANALYTIC THINKING II: TOWARD ANALYTICAL ENGINEERING,” AND LECTURE 4
...............................................................................................................................................................28
10. CHAPTER 9, "EVIDENCE AND PROBABILITIES," AND LECTURE 5..........................................................31
11. CHAPTER 10, "REPRESENTING AND MINING TEXT," AND LECTURE 5..................................................34
12. CHAPTER 12, "OTHER DATA SCIENCE TASKS AND TECHNIQUES," AND LECTURE 5...............................37
13. CHAPTER 13, "DATA SCIENCE AND BUSINESS STRATEGY," AND LECTURE 6..........................................40
14. CHAPTER 14, "CONCLUSION," AND LECTURE 6...................................................................................43
,1. CHAPTER 1, “INTRODUCTION: DATA-ANALYTIC THINKING”, AND LECTURE 1
1. The Core Definitions: Data Science vs. Data-Driven Decision Making
To understand Strategy Analytics, you must distinguish between the activity of analysis and the strategic
approach to decision-making.
Data-Driven Decision-Making (DDD): This refers to the practice of basing decisions on the
analysis of data rather than purely on intuition or experience.
o The Value of DDD: Research by Brynjolfsson et al. shows that firms adopting DDD are
statistically more productive (by 4%–6%) and have higher returns on assets and equity than
firms that rely on intuition.
o Types of Decisions: DDD applies to two main types of decisions:
1. Discoveries: Analyzing data to find new patterns (e.g., Walmart discovering that
strawberry Pop-Tarts sell 7x more before a hurricane).
2. Repetitive Decisions: Improving the accuracy of massive scale, routine decisions
(e.g., MegaTelCo predicting customer churn for millions of accounts).
Data Science: This involves the principles, processes, and techniques for understanding phenomena
via the automated analysis of data. It is the extraction of knowledge.
o Data Science vs. Data Engineering: Data Science focuses on extracting knowledge (the
"science"). Data Engineering focuses on the hardware, software, and pipelines to process
massive amounts of data (the "plumbing," like Hadoop or MongoDB). While Big Data
technologies (Volume, Variety, Velocity) support data science, using them does not
automatically mean you are doing data science.
2. Data as a Strategic Asset: The Capital One Case
A central theme is viewing data not just as a byproduct of business, but as a strategic asset that requires
investment. This is best illustrated by the Signet Bank (Capital One) case study.
The Problem: In the 1990s, banks offered credit cards with uniform pricing because they lacked
data to differentiate customers. They could not identify which customers were profitable and which
were high-risk.
The Strategy: Signet Bank realized that if they could model profitability, they could offer better
terms to good customers ("skim the cream") and avoid bad ones. However, they lacked the data to
build these models because they had never offered varied terms before.
The Solution (Data as Asset): They treated data as an asset to be acquired. They deliberately offered
credit with random terms to random customers. This resulted in immediate financial losses (bad
loans), but these losses were viewed as the cost of data acquisition.
The Result: The data generated allowed them to build superior predictive models for profitability,
leading to the spin-off of Capital One, which became a market leader by tailoring products to specific
customer risk profiles.
Exam Takeaway: You may need to sacrifice short-term profit to generate the data necessary to build a
competitive advantage.
, 3. Fundamental Data Mining Tasks
You must be able to map a business problem to one of the specific data mining tasks.
Classification: Predicting which of a small set of classes an individual belongs to (e.g., "Will this
customer churn?" -> Yes/No).
Regression (Value Estimation): Predicting a numerical value for an individual (e.g., "How much
will this customer use the service?").
Similarity Matching: Identifying similar individuals based on data (e.g., IBM finding companies
similar to their best customers to generate leads).
Clustering: Grouping individuals by similarity without a specific purpose or target variable (e.g.,
"Do our customers form natural groups?").
Co-occurrence Grouping: Finding associations between entities based on transactions (e.g.,
"People who bought X also bought Y").
Profiling: Characterizing typical behavior (e.g., "What is the normal credit card usage for this
segment?"). This is often used for anomaly/fraud detection.
Link Prediction: Predicting connections between data items (e.g., "Since you and Karen have 10
mutual friends, you should be friends").
Causal Modeling: Helping understand what events or actions actually influence others (e.g., "Did
the ad cause the purchase, or would they have bought it anyway?").
4. Supervised vs. Unsupervised Methods
This is a critical technical distinction in Strategy Analytics.
Supervised Learning:
o Definition: There is a specific target variable (outcome) you are trying to predict.
o Requirement: You need labeled historical data where the value of the target is known.
o Examples: Classification (Target = Categorical, e.g., Churn/No Churn) and Regression
(Target = Numerical, e.g., Revenue).
o Evaluation: Can be mathematically evaluated because we can compare predictions to actual
known outcomes.
Unsupervised Learning:
o Definition: There is no target variable. The goal is exploration or pattern finding.
o Examples: Clustering, Profiling, Co-occurrence grouping.
o Evaluation: Harder to evaluate because there is no "correct" answer to compare against.
Exam Tip: If a question asks "Can we find groups of customers who are likely to cancel?", this is
Supervised (Target = Cancel). If it asks "Do our customers fall into natural groups?", this is Unsupervised
(No target).
5. The Data Mining Process (CRISP-DM)
Data science is not a linear software development cycle; it is an exploratory cycle codified by the CRISP-
DM framework.
1. Business Understanding: Defining the problem to be solved. Creativity is essential here to recast
business problems as data science problems.
2. Data Understanding: Estimating the costs and benefits of data sources. Determining if the data
matches the problem (e.g., checking for biases).
3. Data Preparation: Converting data into a tabular format, removing missing values, and preventing
"leaks" (variables that give away the target but won't be available in production).
4. Modeling: Applying data mining techniques (algorithms) to the data to extract patterns.
