In 1872 Aaron Montgomery Ward produced his first catalog and launched the modern era of direct marketing. One hundred thirty-six years later, many direct marketers still rely on Ward’s traditional model to promote their products and services despite the many facets of the unknown that exists with this approach. That’s not to say that direct marketing hasn’t gained sophistication through the years, because it has. Marketers have moved from blanket mailings to using increasingly sophisticated post-campaign analysis to understand which contacts lead to purchases. However, it has long been recognized that not all purchases from customers are a direct result of marketing outreach and therefore are not necessarily a true incremental purchase, i.e., some of the people contacted would have purchased anyway.


Traditional Response Modeling Wastes Money


All marketers struggle with equating marketing dollars spent to wallet share gained with targeted customers. This struggle ensues due to the fact that traditional modeling does not account for the people for which the marketing campaign does not elicit the desired response. That’s right. Not everyone appreciates your hard work and effort. In fact, your direct mail may actually incite a customer to flee!


The following represents the three major problem response categories associated with using traditional response models:


Why target them if they will buy anyway: This problem is simply money wasted targeting people who would buy even if you didn’t target them. These are your loyal customers and customers who planned to purchase the product in advance of receiving the incentive. Why send a coupon for 20 percent off when they were planning to purchase your product anyway?


The highly unlikely to buy issue: Money wasted on people who are highly unlikely to buy is a lost cause, hence the term. These people do not have a purchase pattern that indicates an incremental purchase is on the horizon.


Best left alone segment issue: Imagine this - you actually drive away business by targeting people best left alone. The people in this category do not want to hear from you and in some cases, your direct marketing, may actually push them to never purchase again. We will refer to this group as the “Do Not Disturbs.”


The inability to identify the negative impact a campaign may have on a customer results in wasted marketing dollars as well as customer defections. The bottom line is that the effect of focusing solely on the probability of purchase or traditional response modeling results in every campaign costing more than it needs to, achieving less than it could and annoying some customers unnecessarily.


The Must-Have Customer Segment: The Do Not Disturbs and The Persuadables


There are two groups of people that every marketer must be able to identify before launching a campaign - The Do Not Disturbs and The Persuadables. These groups are the difference between a successful campaign and a less-than-successful campaign. They make all the difference.


The Do Not Disturbs are the group of people for whom the campaign triggers a negative response. Imagine taking all that time and effort working to refine and finesse a marketing campaign only to send customers running in the opposite direction. Unfortunately, traditional response or attrition modeling only focuses on likelihood to respond and cannot identify The Do NotDisturbs.


On the flip side are The Persuadables. Marketers constantly find themselves in pursuit of this group. These are the people who respond in just the way a marketer hopes and makes the job of equating money spent on marketing to the bottom line more straightforward. The group of Persuadables are those that buy (or renew) but would not have done so had they not been recipients of a marketing campaign. Essentially, this is the only group a marketer should want to target because they provide the greatest return on marketing investment (ROMI).


Traditional response models effectively assume that all purchases during or in some period after the campaign are incremental, i.e., would not have happened if the campaign had not been carried out. They also implicitly assume that no purchases are lost as a result of the campaign. History has taught us that is simply not true. So, when these assumptions are not correct, conventional response models will be misleading, and new, more sophisticated uplift models (a.k.a. incremental models) will be likely to perform much better.


Predicting Customer Impact


Uplift modeling predicts the incremental impact – or uplift – a marketing campaign has on a customer and differentiates The Persuadables from The Do Not Disturbs. Uplift modeling measures the variation in the difference between a treated group and a control group. Sometimes discussed using different names – differential response analysis, incremental impact modeling, net modeling, incremental response modeling and "true" response modeling – uplift modeling separates "true" responses to campaigns from purchases that would have been likely to have happen anyway. Uplift modeling determines the increase in the probability that each customer will buy (or stay) when they would otherwise not have done so. You can also see which prospects are likely to ignore the offer (Sure Things and Lost Causes) and which are likely to use it as a trigger to defect (Do Not Disturbs).



Every campaign – whether cross-selling credit cards or trying to prevent attrition for home insurance – creates four distinct response segments as detailed in the diagram


Uplift modeling identifies these critical customer response segments before you run a campaign, so you can target The Persuadables and leave everyone else alone. This helps dramatically increase campaign profitability by allowing you to target fewer people and get an even higher response rate – never mind the additional campaign resources it frees up. In short, you make more money for your company by spending less on direct marketing. This is a concept that seems like an impossibility to some. But as more and more companies adopt and validate uplift modeling, and leading industry analysts recognize its value, word is getting out on the street about the possibilities of this new generation of segmentation modeling.


An uplift model helps to ensure that targeting is used to its maximum effect by directly modeling the difference in outcome (purchase behavior) between a treated group of customers and a comparable control group, and in doing so separates "true" responses from purchases that would in fact have been likely to happen anyway. While it is standard practice to use a control group to measure the overall uplift (sometimes referred to as the net or incremental effect, or the lift associated with a campaign), extending this use of control groups to the modeling process is much more unusual.


Uplift modeling also compares a treated group with a control group and directly models the difference in their attrition behavior. The result is that rather than delivering either attrition likelihood or an estimated value for lost revenue from the customer, uplift modeling delivers either an estimated probability that the intervention will prevent attrition, in other words, an estimate of the value that is likely to be retained as a result of intervention. This is exactly the information that is needed to make optimal targeting decisions for retention.


Real-World Applications of Uplift Modeling: Financial Services


Increasing the length of customer relationships is a growing issue in financial services. The reasons for this are well known: increasing competition combined with regulatory changes to increase customers' willingness to switch.


Uplift modeling has been proven to effectively address cross-sell targeting problems for banks. A typical scenario involves cross-selling activity aimed at increasing product holding. The value of many banking products is high, so that even an increase in product take-up as low as a tenth of a percentage point can provide a ROI for mailings. However, with appropriate targeting, banks can usually achieve between 80 percent and 110 percent of the same incremental sales while reducing mailing volumes by factors ranging from 30 percent to 80 percent. Because the banks in question have themselves attempted to model uplift, they typically have historical data allowing full longitudinal validation of results.



This graph shows an example of one such campaign. Here, the net effect of the campaign was to increase the uptake of the product by a quarter of a percentage point. However, the uplift model shows that over 60 percent of the increase in sales comes from just 10 percent of the targeted population, 90 percent comes from 40 percent of the population and 99 percent comes from 70 percent of the population. Notice also that the bank’s own champion model pro­duced substantially worse results than random targeting - in fact, in this case, reversing the ranking from it would have been much more effective than using its actual output. This suggests that this campaign is being effective in stimulating demand from the very people who tend not to purchase without intervention.


In another example, a global financial services organization had a retention campaign that originally targeted 7,972 from a total population of 32,615. The treatment cost was $5.00 per head and the campaign generated a benefit of $67,786, giving a campaign profit of $27,926 or $3.50 per head, assuming a per head benefit of $208. Using uplift modeling, these figures could be dramatically improved:

  • If you were to target the best 7,972 from the same target population on the basis of the uplift model, the campaign profit achieved would be $150,728 or $18.90 per head. This is a 440 percent improvement on the original targeting and an extra $15.40 incremental revenue per head.
  • If, however, we focus only on those customers where we are having a strong positive effect and maintain adequate control groups, we would target 40 percent of the population. This would deliver a campaign profit of $180,560 or $10.30 per head, by treating 17,530 customers. This is a 194 percent improvement on the original targeting and an extra $6.80 incremental revenue per head.

Use of uplift modeling allows companies to:

  • Avoid contact costs for customers who are not strongly affected by the contact (and for those who may be negatively effected by it);
  • Avoid giving unnecessary incentives and discounts to customers who are likely to purchase regardless;
  • Set mailing volumes on an accurate estimate of the marginal profitability of including customers in campaigns, and;
  • Correctly prioritize those customers for whom the action or offer is maximally effective.

The next generation of marketing modeling is here with the uplift model. And although change is something so many of us resist, it behooves any marketer responsible for a significant direct marketing budget to consider the possibilities the uplift model offers. From saving money to targeting the most profitable customers, the opportunity to create efficiencies within marketing campaigns is here for the taking.

Register or login for access to this item and much more

All Information Management content is archived after seven days.

Community members receive:
  • All recent and archived articles
  • Conference offers and updates
  • A full menu of enewsletter options
  • Web seminars, white papers, ebooks

Don't have an account? Register for Free Unlimited Access