8

Rating ?

Buy book or audiobook

Predictive Analytics
The Power to Predict Who Will Click, Buy, Lie, or Die
Eric Siegel • Wiley © 2016 • 332 pages

Technology / Big Data

Industries / Technology Industry

Take-Aways
• Predictive analytics (PA) is everywhere – in business, manufacturing, health care, government and law
enforcement.
• Prediction analysis can forecast loan defaults, illnesses, car crashes and buying behavior.
• Prediction is imperfect, but it’s far more accurate than simple guesswork.
• In any model, weak data yield bogus results.
• Instead of relying on managers’ experience and intuition, PA mines data for more clinical – and accurate
– decision making.
• Forecasting seeks to make large, sweeping predictions, while PA focuses on individual behaviors: not
how Ohio will vote, but which voters will opt for a particular choice.
• Target combined baby registry data with shoppers’ behavior to predict when customers were pregnant,
provoking a firestorm of criticism.
• The credit score, which dates to the 1940s, was the first use of predictive analytics.
• Decision trees build predictive models and use simple yes-no questions to pinpoint risk.
• The more sensitive data are, the more valuable data become.

1 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST

 Recommendation
Predictive analytics (PA) is a concept that’s both undeniably powerful and potentially creepy. This branch
of computer science combines big data with statistics to foretell what you might buy, how you might vote
or when you might die. Author and predictive analytics guru Eric Siegel is an unabashed cheerleader for
his discipline, and he mostly brushes off the privacy concerns it raises. But that’s not to dismiss his study,
which is engagingly written and elegantly translates dense materials and abstract concepts into easy-to-
read prose. Siegel draws on a number of real-world examples from well-known companies, including Target,
Hewlett-Packard and Chase Bank, and he describes his own experience as an expert consultant on predictive
analytics. The result is an enlightening, plainly written guide. getAbstract recommends his PA overview to
managers and investors seeking insight into a fast-growing corner of the tech economy.

Summary

Big Data and Predictions

Imagine it’s 2022. Predictive analytics (PA) plays a crucial role in your life, starting with your commute to
work. When you get in your car, a predictive model will use biometrics to verify your identity. Spotify will
pick music based on its forecast of your musical tastes.

“Organizations are literally keeping kids in school, keeping the lights on and keeping
crime down with predictive analytics.”

As you drive, a “social techretary” reads the Facebook feeds, CareerBuilder ads and Match responses it
predicts you will like. You’ll receive directions and notifications about traffic congestion from Siri. If you
take your eyes off the road for too long, your car will shake your seat as a warning. If a more severe threat
looms, such as a distracted driver in another vehicle or a child about to dart into traffic, your car will sound a
warning. Your vehicle’s prediction system will scan the engine and transmission and warn you of imminent
mechanical breakdowns.

“Data embody a priceless collection of experience from which to learn.”

These interactions are just the most obvious uses of PA. Other examples operate behind the scenes. You
bought the car with a loan that won approval because you have a good credit score. Your auto insurer
monitors your behavior with sensors that send information about your driving habits to a predictive model.
The insurance company uses this information to set your premium. Your car and your phone predict
cybersecurity threats and fend them off. Government systems forecast such hazards as bridge failures and
manhole explosions.

“PA leads within the growing trend to make decisions more ‘data driven,’ relying less on
one’s ‘gut’ and more on hard, empirical evidence.”

www.getabstract.com
2 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST

 These features may sound futuristic, but they all exist now or are under active development. Smartphones,
smartwatches and web-enabled glasses generate increasing amounts of data. This accumulates due to a
“multiplicative effect” that makes big data and PA ever-more powerful.

Data, Data, Everywhere

Predictive analytics is everywhere. Business, manufacturing, health care, government and law enforcement
deploy it actively. The use of prediction tools can augur success. Top-producing sales teams use PA four
times more than low-performing sales teams. PA is an exploding field. McKinsey & Company reports that
the US job market already faces a shortage of 140,000 analytics specialists.

“The logical flow of a decision tree amounts to a simple computer program, so, in
growing it, the computer is literally programming itself.”

Society is in the throes of “a veritable Big Bang” of big data, the raw material of predictive analysis. Various
mechanisms record every one of your actions – credit applications, Facebook posts, doctor visits and
Amazon transactions – and stores it as a piece of data in a fast-growing warehouse. Each day, data expand
by some 2.5 quintillion bytes. Analysts have deep veins to mine, and organizations that can make sense of
data gain a great advantage.

Applying PA

Companies use PA to find opportunities, such as customers who are likely to buy, and to warn against risks,
such as a possible accident or theft. Examples of how organizations use PA include:

• Buying behavior – Netflix recommends films you’ll like. Hollywood firms analyze screenplays to
forecast ticket sales. Banks and candy makers target their marketing according to consumers’ past
behavior. Utilities forecast the demand for electricity, and Wall Street firms predict stock prices.
• Defections – Which employees will quit, and which customers will jump ship? Companies use PA to
manage their risks from disgruntled workers and clients. Hewlett-Packard rates its 300,000 employees
for “flight risk” – allowing managers to entice workers to stay or to plan for their departure. Mobile
phone companies predict which customers might leave. FedEx forecast customer departures with more
than 65% accuracy. Universities determine which students are at risk of dropping out.
• Accidents and defaults – Insurers use data to predict which motorists are likely to crash. Banks and
collection agencies predict which borrowers are likeliest to default and who among the deadbeats is likely
to pay at last.
• Illness – Hospitals and health insurers increasingly predict who will become hospitalized, how long
those patients will stay in the hospital and how long they’ll live. New tests accurately forecast cancer,
mental illness, and even premature births.
• Crime – Citizens Bank used PA to reduce check fraud by 20%. Hewlett-Packard saved millions by
flagging fraudulent warranty claims. In Chicago and Memphis, police patrol areas where data predict
crimes will occur. The IRS uses PA to pinpoint tax cheats.

www.getabstract.com
3 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST

 Imperfect, Powerful Predictions

Predictive analysis is “technology that learns from experience (data) to predict the future behavior of
individuals in order to drive better decisions.” Predictive analytics is far from perfect, but it still outperforms
an educated guess. For example: Say a company’s direct mail promotion yields a 1% response rate. The
company deploys PA technology to learn which consumers are more likely to respond to its marketing.
Using predictive technology, it targets consumers who show a response rate of 3%. Seen from one
perspective, PA failed miserably – 97% of direct mail pieces wound up in the trash. But PA made the
campaign three times more effective. One mantra of PA states: “A little prediction goes a long way.” One
insurance carrier says using PA to cut its loss ratio by half a percentage point saves nearly $50 million a year.

“As the decision tree becomes bigger and more complex, the predictive performance
continues to increase, but more gradually.”

Before big data, managers made decisions based on their experience, intuition and knowledge. These tools
are valid, but they can be prone to bias. Data provide a more clinical and accurate approach to decision
making. Predictive analytics is its own discipline, a rigorous field relying on computer science and statistics.
PA makes decisions based on evidence, not instinct.

“A decision tree grows upon the rich soil that is data, repeatedly dividing groups of
individuals into subgroups.”

Predictive analytics rely in large part on “collective learning” – an organization compiles data from past
experience and applies the results to future decisions. Used properly, PA can save you from poor decisions.
Through the power of prediction, you could save money by not sending out junk mail to the wrong person
or not making a loan to a risky borrower. PA differs from forecasting, which seeks to make large, sweeping
predictions. PA focuses on individual behaviors. Forecasting tries to guess which presidential candidate will
win Ohio. PA focuses on which individual Ohio voters are likely to vote for a particular candidate.

Target

The retailer Target learned to identify which of its customers were pregnant or were likely to be. Lacking
access to medical records, Target started with another reliable data source: its own baby registry. Expectant
mothers give their due dates when they register at Target for baby gifts. Since not all pregnant women
register with Target, the company took additional steps to predict which customers were pregnant so it could
market to them.

“While prediction itself may be an involved task, it only takes basic arithmetic to
calculate the value realized once prediction is working.”

To build its predictive model, Target combined its baby registry with other data, and looked for patterns. The
retailer learned that pregnant women tended to buy certain products, many of them seemingly unrelated to
babies. The model yielded 30% more customers for marketing targeted to expectant mothers.

www.getabstract.com
4 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST

 “Built upon consumer science and statistics and bolstered by devoted conferences and
university degree programs, PA has emerged as its own discipline.”

Prediction professionals would say that Target’s pregnancy data investigation was a clear, unambiguous
success. The company mined its own information to boost sales. But when the rest of the world learned of
the practice, Target’s use of data didn’t seem so benign. In February 2012, The New York Times Magazine
published a piece by reporter and author Charles Duhigg, “How Companies Learn Your Secrets.” Duhigg
implied that PA was an invasion of privacy and allowed predatory companies to manipulate defenseless
consumers.

“An organization that doesn’t leverage its data in this way is like a person with a
photographic memory who never bothers to think.”

In Duhigg’s telling, a man discovered his teen daughter was pregnant by reading Target’s marketing
messages to her. The Target example became part of Duhigg’s best-selling book The Power of Habit, and
ignited a media firestorm about the apparent dark side of predictive analytics – never mind that Duhigg
didn’t prove that Target exposed the anonymous teen. When data analysis divines a personal detail as
intimate as pregnancy, questions about privacy follow.

“A hazy view of what’s to come outperforms complete darkness by a landslide.”

The more sensitive data are, the more valuable data become. Sometimes, the results of PA are embarrassing.
If you recently bought a medication for gas pain at Walgreens, on your next visit you might get a coupon for
gas pills. Consumers dislike data mining. A quarter of Facebook profiles contain bogus personal information
entered by users trying to confuse big data.

How Machines Learn

Predictive analytics relies on machine learning. Predictive models start with “training data” – basic
information such as Target’s baby registry list – and then it learns how that data foretell future behavior. As
analysts introduce more data, the models learn to draw connections. In any model, weak data yield bogus
results.

“Several mounting ingredients promise to spread prediction even more pervasively:

bigger data, better computers, wider familiarity and advancing science.”

In the case of Hewlett-Packard and its employee turnover or flight scores, HP compiled two years of
employee information. The data included salaries, job ratings and job assignments, as well as whether
the employee was still with HP. The company sought to predict which employees were likely to leave.
From the employer’s standpoint, this use of data makes perfect sense. HP can be forewarned of an
impending departure or try to persuade a valued employee to stay. But the scores could make an employee
uncomfortable. What if a loyal worker feels unfairly branded as a flight risk? Recognizing the sensitivities
that a loyalty rating might provoke, HP employs a “handle with care” policy for its flight risk scores. Only
a few managers can see the scores, and the reports identifying individuals by numbers that managers
must unscramble, not by names. At HP, higher salaries and more frequent job rotations correlated with

www.getabstract.com
5 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST

 longer tenures. However, frequent promotions did little to improve loyalty in HP’s sales compensation
team. Promoted sales compensation workers were more likely to leave – apparently a reaction to relatively
small raises. HP’s flight risk scores appeared to generate as much as $300 million in savings by avoiding or
mitigating employee turnover.

Credit Scores

The credit score debuted in 1941 and helped validate the concept of predictive analytics. After credit scores
proved to be effective in reducing risk in consumer lending, banks broadened the practice to loan portfolios.
For instance, after a series of mergers, Chase Bank in 1996 found itself with a massive portfolio of millions
of mortgages. To manage these loans, Chase worked with the concept of “microrisk” and “macrorisk.” An
individual loan was a microrisk. A default or prepayment of a single mortgage wouldn’t affect the megabank.
But having a large number of smaller loans go bad would constitute a macrorisk that could threaten the
bank’s profitability.

“Decision Trees”

Decision trees build predictive models and pinpoint risk by asking simple yes-or-no questions. In one
example, Chase analyzed the risk of loan prepayment (which reduces the interest earned). It found that
19.2% of borrowers with a mortgage interest rate higher than 7.94% would prepay, but for mortgages
with rates less than 7.94%, the prepayment risk was just 3.8%. So, Chase started its decision tree with the
question: Is the rate less than 7.94%? The answer leads to a question about the borrower’s income and
another question about the size of the mortgage. Chase found that the risk of prepayment soared to 36%
for high-rate mortgages with balances of more than $182,926. The more data decision trees use, the more
powerful they become. Chase found, for instance, that if the mortgage was for more than $67,750 but less
than $182,926, and the interest rate was 8.69% or higher, and the loan-to-value ratio was less than 87.4%,
the borrower would prepay in 25.6% of cases. At some point, decision trees reach a point of diminishing
returns. They grow only marginally more accurate as they become increasingly more complex.

“Microrisks matter. Left unchecked, they threaten to snowball. Our best bet is to learn to
predict.”

Decision trees are “simple, elegant and precise” and “practically mathless.” The data need not be numbers
such as loan balances and interest rates. For example, take a decision tree that predicts the votes of US
Supreme Court justices. The model used several hundred past votes to predict each justice’s future rulings. It
correctly foretold 75% of Supreme Court votes. By contrast, human legal experts were right only 59% of the
time.

www.getabstract.com
6 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST

 About the Author
Former Columbia University professor Eric Siegel is the founder of Predictive Analytics World and
executive editor of The Predictive Analytics Times.

Did you like this summary?

Buy book or audiobook
http://getab.li/27852

This document is restricted to the personal use of OSCAR ALVARADO (o.alvarado@dhl.com)

getAbstract maintains complete editorial responsibility for all parts of this review. All rights reserved. No part of this review may be reproduced or
transmitted in any form or by any means – electronic, photocopying or otherwise – without prior written permission of getAbstract AG (Switzerland).

7 of 7

LoginContext[cu=8295682,ssoId=4623603,asp=395,subs=8,free=0,lo=en,co=US] 2024-07-03 18:06:35 CEST