Churn Baby Churn

Before this change, Telenor had been successfully applying the industry's leading technique to hold on to its cell phone subscribers—a technique that applies predictive analytics (PA):

PA Application: Customer Retention with Churn Modeling

1. What's predicted: Which customers will leave.
2. What's done about it: Retention efforts target at-risk customers.

Churn modeling may be the hottest marketing application of PA, and for good reason. Any seasoned executive will tell you retention is all-important because it's usually cheaper to convince a customer to stay than to acquire a new one.

Picture customer turnover as air flowing into and out of a balloon:

Retaining more customers is akin to clamping down on the nozzle on the right. Lessening the rate of loss just a bit, the balloon blossoms, magnifying its rate of expansion—that is, the growth rate of the company's customer base. This growth is the raison d'être of business.

Prediction and proaction are musts. Persuading someone to stay often sets a mobile carrier back a free phone or a hefty discount. A company must target this generosity where it's needed: those customers predicted to leave. Like most major cell phone carriers, Telenor had been enjoying a clear win with churn modeling.¹

What could possibly go wrong?

Sleeping Dogs

If I leave here tomorrow

Would you still remember me?

For I must be traveling on, now

'Cause there's too many places I've got to see.

—From “Free Bird” by Lynyrd Skynyrd

Imagine you received an alluring brochure from your cell phone company that says:

Tantalized? Imagining a higher-tech toy in your pocket?

Now imagine you are newly emancipated, recently granted the liberty to take your phone number with you to another carrier. You've been aching to change to another carrier to join your friends who say they love it over there. In fact, your provider may have sent you this offer only because it predicted your likely departure.

Big mistake. The company just reminded you that your contracted commitment is ending and you're free to defect.

Contacting you backfired, increasing instead of decreasing the chance you'll leave. If you are a sleeping dog, they just failed to let you lie.

Bad news piled on. While already struggling against rising rates of defection, Eva and her colleagues at Telenor detected this backfiring of their efforts to retain, a detrimental occurrence that was now happening more often. More customers were being inadvertently turned away, triggered to leave when they otherwise, if not contacted, might have stayed. It was no longer business as usual.

A New Thing to Predict

You didn't have to be so nice; I would have liked you anyway.

—The Lovin' Spoonful, 1965

D'oh!

—Homer Simpson

This newly dominant phenomenon brought up for Telenor the question of what PA should be used to predict in the first place. Beyond predicting departure, must a company secondarily predict how customers will respond when contacted? Must we predict the more complicated, two-part question, “Who is leaving but would stay if we contacted them?” This sounds pretty convoluted. To do so, it seems like we'd need data tracking when people change their minds!

This question of integrating a secondary prediction also pertains to another killer app of PA, the utterly fundamental targeting of marketing:

PA Application: Targeted Marketing with Response Modeling

1. What's predicted: Which customers will purchase if contacted.
2. What's done about it: Contact those customers who are more likely to do so.

Despite response modeling's esteemed status as the most established business application of PA (see the 12 examples listed in this book's Central Table 2), it falls severely short because it predicts the outcome for those we do contact, but not for those left uncontacted. Assume we have contacted these individuals:

If the dark gray individuals made a purchase, we may proceed with patting ourselves on the back. We must have done a great job of targeting by way of astute predictions about who would buy if contacted, since so many actually did so—relative to how direct marketing often goes, achieving response rates of a few percent, 1 percent, or even less.

One simple question jolts the most senior PA expert out of a stupor: Which of the dark gray individuals would have purchased anyway, even if we hadn't contacted them? In some cases, up to half of them—or even more—are so prone to purchasing, they would have either way.

Even an analytics practitioner with decades of experience tweaking predictive models can be floored and flabbergasted by this. She wonders to herself, “Have I been predicting the wrong thing the whole time?” Another bonks himself on the head, groaning, “Why didn't I ever think of that?” Analytics labs echo with the inevitable Homer Simpson exclamation, “D'oh!”

Let's step back and look logically at an organization's intentions:

• The company wants customers to stay and to buy.
• The company does not intend to force customers (they have free will).
• Therefore, the company needs to convince customers—to influence, to persuade.

If persuasion is what matters, shouldn't that be what's predicted? Let's try that on for size.

Prediction goal: Will the marketing brochure persuade the customer?

Mission accomplished. This meets the company's goals with just one predictive question, integrating within it both whether the customer will do what's desired and whether it's a good idea to contact the customer.

Predicting impact impacts prediction. PA shifts substantially, from predicting a behavior to predicting influence on behavior.

Predicting influence promises to boost PA's value, since an organization doesn't just want to know what individuals will do—it wants to know what it can do about it. This makes predictive scores actionable.

I know I asked this earlier but, what could possibly go wrong?

Eye Can't See It

Houston, we have another problem.

How can you know something happened if you didn't see it? Take a look at this possible instance of influence:

1. The individual perceives the sales brochure.
2. Something happens inside the brain.
3. The individual buys the product.

Is it safe to assume influence took place? How do we know the brochure made a difference? Perhaps the individual would have purchased anyway.

The brain's a black box into which we cannot peek. Even if we were conducting neuroscience, it's not clear if and when that field of science will progress far enough to detect when one changes one's mind (and even if it could, we'd need brain readings from each consumer to employ it!).

Introspection doesn't work, either. You cannot always report on how your own decision making took place. You just can't be certain what made a difference, whether your friend, client, sister, or even you yourself would have made a purchase if circumstances had been different.

To observe influence, we'd need to detect causality: Did the brochure cause the individual to purchase? As explored in Chapter 3, our knowledge about causality is limited. To truly know causality would be to fully understand how things in the world affect one another, with all the detail involved, the chain reactions that lead one event to result in another. This is the domain of physics, chemistry, and other sciences. It's How the World Works. Ultimately, science tells us only a limited amount.

Therefore, influence cannot be observed. We can never witness an individual case of persuasion with complete certainty.

How, then, could we ever predict it?

Perceiving Persuasion

No man ever steps in the same river twice.

—Heraclitus

Good grief. The most valuable thing to predict can't even be detected in the first place.

The desire to influence drives every move we make. As organizations or individuals, out of self-interest or altruistically, almost everything we do is meant to produce a desired effect, including:

• Send a brochure to a customer (or voter).
• Prescribe a medication to a patient.
• Provide social benefits intended to foster self-sufficiency.

Each action risks backfiring: The customer cancels, the patient suffers an adverse reaction, or the beneficiary becomes dependent on assistance. So we make choices not only to pursue what will work, but also to avoid what would do more harm than good.

In one arena in particular, do we feel the pangs of misstep and failure: dating. In courtship, you are both the director of marketing and the product. You're not in the restaurant for food—rather, it is a sales call. Here are some tips and pointers to persuade. Don't be overly assertive, too frequently contacting your prospect. Yet don't remain overly passive, risking that a competitor will swoop in and steal your thunder. Try to predict what you think is the right message, and avoid communicating the wrong thing.

In the movie Groundhog Day, our hero Bill Murray acquires a kind of superpower: the coveted ability to perceive influence. Stuck in a magical loop, reliving the same dull day over and over, he faces a humbling sort of purgatory, apparently designed to address the character's flamboyant narcissism. He cannot escape, and he becomes despondent.

Things turn around for Bill when he recognizes that his plight in fact endows him with the ability to test different marketing treatments on the same subject under exactly the same circumstances—and then observe the outcome. Desperate to win over the apple of his eye (Andie MacDowell) and immune to the fallout and crush of failure, he endeavors in endless trial and error to eventually learn just the right way to woo her.

Only in this wonderful fantasy can we see with certainty the difference each choice makes. That's life. You never know for sure whether you made the optimal choice about anything. Should I have admitted I love the Bee Gees? Should we have sent that brochure? Would the other surgical treatment have gone better? Woulda, coulda, shoulda.

In real life, there are no do-overs, so our only recourse is to predict beforehand as well as possible what will work. But, in real life, what's real? If we can't observe influence, how do we know it ever really happens at all?

Persuasive Choices

Think before you speak.

Even in dating, there's science to persuasion. Dating website OkCupid showed that messages initiating first contact that include the word awesome are more than twice as likely to elicit a response as those with sexy. Howdy is better than hey. Band does better than literature and video games (go figure).

Psychology professor Robert Cialdini persuaded people to commit less crime, and proved it worked. Visitors regularly steal a precious resource from Arizona's Petrified Forest National Park: chunks of petrified wood. Cialdini measured the result of posting the following sign:

With that sign in place, the rate of theft was 1.67 percent. Next he tested another message that more strongly emphasizes the negative effect of theft:

You might expect that would further reduce theft, but it backfired. This message has the effect of destigmatizing theft, since it implies the act is common—“Everybody does it.” Possibly for that reason, it resulted in more than four times as much theft as the first sign, 7.92 percent. Regardless of the psychological interpretation and whether the result is a surprise, persuasion has been proven. We can safely conclude that relaying the first message rather than the second influences people to steal less. Similar effects have been shown in the persuasion of hotel room towel recycling and decreasing home energy usage, as explored in Cialdini's coauthored book, Yes! 50 Scientifically Proven Ways to Be Persuasive.²

These studies prove influence takes place across a group but ascertain nothing about any one individual, so the choice of message still cannot be individually selected according to what's most likely to influence each person.

In the field of medicine, most clinical studies do this same thing—compare two treatments and see which tends to work better overall. For knee surgery after a ski accident, I had to select a graft source from which to reconstruct my busted anterior cruciate ligament (ACL, the knee's central ligament—previously known to me as the Association for Computational Linguists). I based my decision on a study that showed subsequent knee walking was rated “difficult or impossible” by twice as many patients who donated their own patellar tissue rather than hamstring tissue.³

It's good, but it's not personalized. I can never know if my choice for knee surgery was the best for my particular case (although my knee does seem great now). The same holds true for any treatment decision based on such studies, which provide only a one-size-fits-all result. We're left with uncertainty for each individual patient. If you take a pill and your headache goes away, you can't know for sure that the medicine worked; maybe your headache would have stopped anyway.

More generally, if you prevent something bad, how can you be sure it was ever going to happen in the first place?

Business Stimulus and Business Response

Many of your everyday clicks contribute to the Web's constant testing of how to improve overall persuasiveness. Google has compared 41 shades of blue to see which elicits more clicks. Websites serve the ads that get clicked the most and run random AB tests to compare which Web page design and content lead to the most buying. Facebook conducts controlled experiments to see how changes to the rules driving which friends' posts get displayed influence your engagement and usage of their website (see Central Table 1).

I tested titles for this book, following in the footsteps of SuperCrunchers and The 4-Hour Workweek. Placed as ads on Google Adwords, Predictive Analytics, when displayed on tens of thousands of screens of unsuspecting experimental subjects across the country, was clicked almost twice as often as Geek Prophecies and also beat out I Knew You Were Going to Do That and Clairvoyant Computers, plus six other book titles that I also entered into this contest. It was convenient that the field's very name came out as the top contender, an unquestionably fitting title for this book.

In both medicine and marketing, this scheme to test treatments reveals the impact of selecting one outward action over another—but only as a trend across the group of subjects as a whole. After this sort of experiment, the best an organization can do is run with the one most effective treatment, applying it uniformly for all individuals.

In this practice, the organization is employing a blunt instrument. Looking back, we still don't know for whom the treatment was truly effective. Looking forward, we still don't know how to make personalized choices for each individual.

The Quantum Human

Here's the thing about the future. Every time you look at it, it changes. Because you looked at it.

—Nicolas Cage's clairvoyant in Next

Heisenberg might have slept here.

—Anonymous

As in quantum physics, some things are unknowable. Although you may protest being reduced to a quantum particle, there's a powerful analogy to be drawn between the uncertainty about influence on an individual and Heisenberg's uncertainty principle. This principle states that we can't know everything about a particle—for example, both its position and speed. It's a trade-off. The more precisely you measure one, the less precisely you can measure the other.

Likewise, we can't know everything about a human. In particular, we can't know both things that we'd need to know in order to conclude that a person could be influenced. For example:

1. Will Bill purchase if we send him a brochure?
2. Will Bill purchase if we don't send him a brochure?

If we did know the answer to both, we'd readily know this most desired fact about Bill—whether he's influenceable. In some cases, the answers to the two questions disagree, such as:

The answer to (1) is “Yes”—Bill receives a brochure and then purchases.

The answer to (2) is “No”—Bill does not receive a brochure and does not purchase.

In this case, we would conclude that the choice of treatment does have an influential effect on Bill; he is persuadable.

In other cases, the answers to the questions agree, such as:

The answer to (1) is “Yes”—Bill receives a brochure and then purchases.

The answer to (2) is also “Yes”—Bill does not receive a brochure but then purchases anyway.

In this case, we conclude the choice of treatment has no influence; he would buy either way. This type of customer is called a sure thing.

Other scenarios exist. Sometimes a brochure backfires and adversely influences a customer who would otherwise buy not to.

But this is a fantasy—we can't know the answer to both questions. We can find out (1) by sending Bill a brochure. We can find out (2) by not sending him a brochure. But we can't both contact and not contact Bill. We can't administer medicine and not administer medicine. We can't try two different forms of surgery at once. In general, you can't test an individual with both treatments.

This uncertainty leaves us with philosophical struggles akin to those of quantum physics. Given that we could never know both, does a particle ever really have both a true position and a true speed? Similarly, do answers to both of the previous questions about a person truly exist? Answering one renders the other purely hypothetical. It's like the tree falling in the forest with no one to perceive the sound, which becomes only theoretical. This most fundamental status of a human as influenceable or not influenceable holds only as an ethereal concept. It's only observable in aggregate across a group, never established for any one person. Does the quality of influenceability exist only in the context of a group, emergently, defying true definition for any single individual? If influenceable people do walk among us, you can never be certain who they are.

The quantum human—is he or she influenceable?

This unknowability equates the past and the future. We don't know whether a person was influenced, and we don't know whether the person could be influenced—whether he or she is influenceable. It's kind of a refreshing change that prediction is no more difficult than retrospection, that tomorrow presents no greater a challenge than yesterday. Both previous and forthcoming influence can only at best be estimated. Clearly, the future is the more valuable one to estimate. If we can know how likely each person is to be influenced, we can drive decisions, treating each individual accordingly.

But how can you predictively model influence? That is, how could you train a predictive model when there are no learning examples—no individual known cases—of the thing we want to predict?

Predicting Influence with Uplift Modeling

A model that predicts influence will be a predictive model like any other:

Like all the models we've covered in this book, it takes characteristics of the individual as input and provides a predictive score as output.

But it will be a special case of predictive models. Instead of predicting an outright behavior, we need a model that scores according to the likelihood an individual's behavior will be influenced. We need an uplift model:

1. Uplift model—A predictive model that predicts the influence on an individual's behavior that results from applying one treatment over another.⁴

The uplift score answers the question, “How much more likely is this treatment to generate the desired outcome than the alternative treatment?” It guides an organization's choice of treatment or action, what to do or say to each individual.⁵ The secondary treatment can be the passive action of a control set—for example, make no marketing contact or administer a placebo instead of the trial drug—in which case an uplift model effectively decides whether or not to treat.

How do you learn about something you can't see? We never have at our disposal learning examples of the very thing we want to predict: influenceable individuals. We don't have the usual training data from which to directly learn.

To do the seemingly impossible, uplift modeling needs a clever work-around. To see how it works, let's explore a detailed example from U.S. Bank.

Banking on Influence

U.S. Bank Assistant Vice President Michael Grundhoefer isn't satisfied with good. In the mid-1990s, the bank's direct marketing efforts to sell financial products such as lines of credit fared well. Most mail campaigns turned a satisfactory profit. Michael, who headed up the analytics behind many of these campaigns, kept a keen eye on the underlying response models and how they could be improved.

Companies often misinterpret marketing campaign results. Here's where they go terribly wrong: They look at the list of customers contacted and ask, “How many responded?” That's the response rate. One of the original inventors of uplift modeling, Nicholas Radcliffe (now an independent consultant and sometimes visiting professor in Edinburgh), drew a cartoon about that measure's drawbacks:

Cartoon reproduced with permission.

The response rate completely overlooks how many would buy anyway, even if not contacted. Some products just fly off the shelves and sell themselves. For business, that's a good thing—but if so, it's important not to credit the marketing. You could be wasting dollars and chopping down trees to send mail that isn't actually helping.

Just as with medicine, marketing's success—or lack thereof—is revealed by comparing to a control set, a group of individuals suppressed from the treatment (or administered a placebo, in the case of medicine). Therefore, we need to collect two sets of data:

If the treated customers buy more than the control customers, we know the campaign successfully persuades. This proves some individuals were influenced, but, as usual, we don't know which.

Predicting the Wrong Thing

If you come to a fork in the road, take it.

—Yogi Berra

To target the marketing campaigns, Michael and his team at U.S. Bank were employing the industry standard: response models, which predict who will buy if contacted. That's not the same thing as predicting who will buy because they were contacted; it does not predict influence. Compared to a control set, Michael showed the campaigns were successful, turning a profit. But he knew the targeting would be more effective if only there were a way to predict which customers would be persuaded by the marketing collateral.

Standard response models predict the wrong thing and are in fact falsely named. Response models don't predict response caused by contact; they predict buying in light of contact. But predicting for whom contact will be the cause of buying is more pertinent than predicting buying in general. Knowing who your “good” customers are—the ones who will buy more—may be nice to know, but it takes second place.⁶

For some projects, conventional response models have it backward. By aiming to increase response rate, they complacently focus on the metric that's easiest to measure. As former U.S. Secretary of Defense Robert McNamara said, “We have to find a way of making the important measurable, instead of making the measurable important.” A standard response model will gladly target customers who would buy anyway, doing little to address how much junk mail we as consumers receive. Instead, it's only a small sliver of persuadable customers who are actually worth mailing to, if we can identify them.

Standard response modeling predicts:

1. Will the customer buy if contacted?

   Uplift modeling changes everything by adding just one word:

2. Will the customer buy only if contacted?

Although the second question may appear simple, it answers the composite of two questions: “Will the customer buy if contacted and not buy otherwise?” This two-in-one query homes in on the difference that will result from one treatment over another. It's the same as asking, “Would contacting the customer influence him or her to buy?”

Response Uplift Modeling

Weigh your options.

By addressing a composite of two questions, each individual belongs in one of four conceptual segments that distinguish along two dimensions:

Conceptual response segments. The lower-right segment is targeted with uplift modeling.⁷

This quad first distinguishes from top to bottom which customers will buy in light of marketing contact, which is the job of conventional response modeling. But then it further distinguishes along a second dimension: Which customers will make a purchase even if not contacted?

Michael at U.S. Bank wanted to target the lower-right quadrant, those worthy of investing the cost to contact. These persuadables won't buy if not contacted, but will buy if they are. These are the individuals an uplift model aims to flag with the affirmative prediction.

PA Application: Targeted Marketing with Response Uplift Modeling

1. What's predicted: Which customers will be persuaded to buy.
2. What's done about it: Target persuadable customers.

An uplift model provides the opportunity to reduce costs and unnecessary mail in comparison to a traditional response model. This is achieved by suppressing from the contact list those customers in the lower-left quadrant, the so-called sure things who will buy either way.

The Mechanics of Uplift Modeling

Uplift modeling operates simultaneously on two data sets—both the treated set and the control set—learning from them both:

Two training sets are used together to develop an uplift model.

To learn to distinguish influenceables—those for whom the choice of treatment makes a difference—uplift modeling learns from both customers who were contacted and others who weren't. Processing two data sets represents a significant paradigm shift after decades of predictive modeling and machine learning research almost entirely focused on tweaking a modeling process that operates across a single data set.

Starting first with a single-variable example, we can see that it is possible to predict uplift by comparing behavior across the two data sets:

Net weight of evidence (NWOE, a measure of uplift) varies by a customer's number of open revolving accounts. Graph courtesy of Kim Larsen.

This fictional but typical example of a financial institution's direct-marketing results illustrates that mildly engaged customers are hot, readily persuadable by direct mail. The vertical axis represents net weight of evidence (NWOE), a measure of uplift, and the horizontal axis represents the number of open revolving accounts the customer already holds. In this case, it turns out that customers in the middle region, who don't already hold too many or too few open revolving accounts, will be more likely to be persuaded by direct mail.

Less engaged customers on the left are unmoved—whether they were already destined to open more accounts or not, their plans don't change if contacted. This includes both sure things and lost causes—either way, it isn't worth contacting them.

Avoid at all costs contacting customers on the right—they are “do-not-disturbs.” Contacting these individuals, who already hold a good number of accounts, actually decreases the chance they'll buy. The curve dips down into negative numbers—a veritable downlift. The explanation may be that customers with many accounts are already so engaged that they are more sensitive to, aware of, and annoyed by what they consider to be unnecessary marketing contact. An alternative explanation is that customers who have already accumulated so many credit accounts are susceptible to impulse buys (e.g., when they come into a bank branch), but when contacted at home will be prone to respond by considering the decision more intently and researching competing products online.

This shows the power of one variable. How can we leverage PA's true potential by considering multiple variables, as with the predictive models of Chapter 4? Let's turn back to Michael's story for a detailed example.

How Uplift Modeling Works

Despite their marketing successes, Michael at U.S. Bank had a nagging feeling things could be better. Unlike many marketers, he was aware of the difference between a campaign's response rate and the sales generated by it. Inspecting reports, he could see the response models were less than ideal. He tried out some good ideas of his own to attempt to model persuasion, which provided preliminary yet inconsistent and unstable success.

One time, Michael noted failure for a certain group within a direct mail campaign selling a home-equity line of credit to existing customers. For that group, the campaign not only failed to cover its own printing and mailing costs, it in fact had the detrimental effect of decreasing sales, a slight downlift overall.

Michael was beginning to collaborate with a small company called Quadstone (now Pitney Bowes Software) that provided a new commercial approach to uplift modeling. The system could derive marketing segments that reveal persuadable customers, such as:⁸

A segment of persuadable individuals.

This is not your usual marketing segment. It doesn't designate customers more likely to buy. It doesn't designate customers less likely to buy. It is customers more likely to be influenced by marketing contact. The difference marketing makes for this segment can be calculated only by seeing how its purchase rate differs between the treated and control sets:⁹

Purchase rates of the persuadable segment described above differ, depending on whether marketing contact is received.

Success! Within this segment, the direct mail elicits more responses from customers who were contacted (the treated set) than those not contacted (the control set). By automatically deriving its defining characteristics, uplift modeling has discovered a segment of customers for which this direct mail campaign succeeds after all.

The uplift modeling method that discovers such segments is an expansion of decision trees (see Chapter 4) called uplift trees. Normal decision trees strive to identify segments extreme in their response rates—many responses or few responses. Uplift trees use variables to mechanically “segment down” in the same way, but seek to find segments extreme in the difference treatment makes—segments that are particularly influenceable. A single uplift tree is composed of a number of segments such as the one shown above.¹⁰

For U.S. Bank, response uplift modeling delivered an unprecedented boost, increasing the marketing campaign's return on investment (ROI) by a factor of five in comparison with standard response model targeting. This win resulted from reducing both the amount of direct mail that commanded no impact (sent to lost causes or sure things) and the amount that instigated an adverse response (sent to sleeping dogs, aka do-not-disturbs).

Uplift practitioners at Fidelity Investments also see the light: Spend less, earn more. By avoiding sure things and do-not-disturbs, “Uplift modeling empowers your organization to capture more than 100 percent of responses by contacting less than 100 percent of the target population,” says Kathleen Kane, Fidelity's principal decision scientist.

The Persuasion Effect

Uplift modeling conquers the imperceivable—influence—by newly combining two well-trodden, previously separate paradigms:

1. comparing treated and control results; and
2. predictive modeling (machine learning, statistical regression, etc.).

Only by cleverly combining these two practices does the newfound ability to predict persuasion for each individual become possible. I call this The Persuasion Effect:

If you haven't already figured it out, this answers the riddle posed at the beginning of this chapter. Being influenced is the thing that often happens to you that cannot be witnessed and that you can't even be sure has happened afterward—but that can be predicted in advance. In this way, PA transcends human perception.

Influence across Industries

Uplift modeling applies everywhere: marketing, credit risk, electoral politics, sociology, and healthcare. The intent to influence is common to almost all organizations, so The Persuasion Effect is put into play across industry sectors.

This chapter covers in detail the first two areas on marketing in the table above, as well as a case study in electoral politics (in the sidebar about Obama's 2012 presidential campaign at the end of this chapter). Here's a bit more about the rest of them (note that for some of these application areas, no public case studies or proofs of concept yet exist—uplift modeling is an emerging technology).

Content and channel selection. Uplift modeling selects for each user the ad, offer, content, product, or channel of contact (phone, e-mail, etc.) most likely to elicit a response. In these cases, there is no passive option and therefore no control set—both data sets test an active treatment.

Dynamic pricing and collections. As for any decision, a certain risk is faced for each treatment option when pricing: The higher price may turn a customer away, but the lower price (or deeper discount or write-off, for collections) unnecessarily sacrifices revenue if the customer would have been willing to pay more.

Credit risk. The balance between risk and upside profitability for each debtor is influenced by both the credit limit and the APR offered. Raising one or both may result in higher revenue in the form of interest payments, but may also increase the chance of the debtor defaulting on payments and an ensuing write-off.

Electoral politics. As a resident of California, I see few if any ads for presidential campaigns—the state is a lock; depending on your political affiliation, it could be viewed as either a sure thing or a lost cause. Just as so-called swing clients (influenceables) are potentially persuaded by marketing contact, the same benefit is gained where this term originates: political campaigns that target swing voters. The constituents with the most potential to be influenced by campaign contact are worth the cost of contact. Analogously, only the swing states that could conceivably be persuaded as a whole are worth expending great campaign resources. For more on elections and uplift modeling, see this chapter's sidebar, “Beyond Swing Voters: How Persuasion Modeling Revolutionized Political Campaigns for Obama and Beyond.”

Sociology: targeting social programs. Speaking of politics, here is a concept that could change everything. Social programs such as educational and occupational support endure scrutiny as possibly more frequently awarded to go-getters who would have succeeded anyway. For certain other beneficiaries, skeptics ask, does support backfire, leaving them more dependent rather than more self-sufficient? Only by predicting how a program will influence the outcome for each individual prospect can programs be targeted in a way that addresses these questions. In so doing, might such scientifically based, individualized economic policies help resolve the crippling government deadlock that results from the opposing fiscal ideologies currently held by conservative and liberal policymakers?

Personalized medicine. While one medical treatment may deliver better results on average than another, this one-size-fits-all approach commonly implemented by clinical studies means treatment decisions that help many may in fact hurt some. In this way, healthcare decisions backfire on occasion, exerting influence opposite to that intended: They hurt or kill—although they kill fewer than following no clinical studies at all. Personalized medicine aims to predict which treatment is best suited for each patient, employing analytical methods to predict treatment impact (i.e., medical influence) similar to the uplift modeling techniques used for marketing treatment decisions. For example, to drive beta-blocker treatment decisions for heart failure, Harvard University researchers “use two independent data sets to construct a systematic, subject-specific treatment selection procedure.” A certain HIV treatment is shown to be more effective for younger children. Treatments for various cancers are targeted by genetic markers—a trend so significant the Food and Drug Administration is increasingly requiring for new pharmaceuticals, as The New York Times puts it, “a companion test that could reliably detect the [genetic] mutation so that the drug could be given to the patients it is intended to help,” those “most likely to benefit.”

Immobilizing Mobile Customers

It wasn't long after phone number portability came, raising a hailstorm in the telecommunications industry, that Quadstone spoke with Eva at Telenor about the new uplift modeling technique. It was a revelation. Eva had already confirmed that Telenor's retention efforts triggered some customers to leave rather than persuading them to stay, but she wasn't aware of any established technique to address the issue. The timing was fortuitous, as Quadstone was just starting out, seeking its first few clients to prove uplift modeling's value.

PA Application: Customer Retention with Churn Uplift Modeling

1. What's predicted: Which customers can be persuaded to stay.
2. What's done about it: Retention efforts target persuadable customers.

Customers can be as easily scared away as a skittish bunny. Traditional churn models often inadvertently frighten these rabbits, since customers most likely to leave are often those most easy to trigger—sleeping dogs easy to wake up. This includes, for example, the health club member who never gets to the gym and the Netflix subscriber who rarely trades in each DVD movie rental—both just need a reminder before they get around to canceling (it would be more ideal to reengage them). Someone once told me that, when he received an offer to extend his camera's warranty, it reminded him that coverage was soon ending. He promptly put his camera into the microwave to break it so he could return it. It would inevitably be more cost-effective to avoid triggering such criminal activity than to prosecute for it after the fact.

Prompting a cell phone customer to leave can be especially costly because it may trigger a social domino effect: People tend to stick with the same wireless carrier as their friends. One major North American carrier showed that a customer is seven times more likely to cancel if someone in the person's calling network cancels.

For Telenor, churn uplift modeling delivered an astonishing boost to the effectiveness of its retention initiatives: The ROI increased by a factor of 11, in comparison with its prior, established practice of targeting with standard churn models. This came from decreasing the number of sleeping dog customers the company had been inadvertently waking, and secondarily from reducing the total number of mail pieces sent—like U.S. Bank, Telenor got more for less.

Figure permission of Pitney Bowes Software.

For the international mobile carrier, which serves tens of millions of cell phone subscribers across 11 markets, this was a huge win. Beyond addressing the new business challenges that came of phone number portability, it alleviated the systematic “sleeping dog” problem inherent to churn modeling, one Telenor likely had suffered from all along. Even when there's a net benefit from marketing, offers could be triggering some customers to leave who would have otherwise stayed.

For Eva, who has since been promoted to head of customer analytics, and for the rest of the world, this only marks the beginning of the emerging practice of inducing influence and predicting persuasion.