By using website you agree to our use of cookies as described in our cookie policy. Learn More


Why Your BI Model Should Include an Uplift Model

Instead of forecasting whether a prospect will take an action once they're contacted, persuasion modeling attempts to measure whether customers will take an action only if they're contacted.

By Zach Watson, TechnologyAdvice

After the dust settled and the furor of the 2012 presidential election subsided, a consistent narrative surrounding Barrack Obama's victory began to take shape: the newly elected president had used data to predict the behavior of swing votes. Armed with a team of data scientists, Obama had conquered his opponents through an unprecedented application of mathematical equations.

Though the subsequent fanfare may have distorted the actual effectiveness of the Obama team's data strategy, one very important experiment was conducted during the campaign: a telemarketing campaign that measured "persuadability."

Uplift modeling -- referred to as persuasion modeling by the Obama campaign -- measures the receptiveness of your audience to being persuaded. In other words, if you communicate with these prospects, does your message actually change their behavior?

Traditional response-marketing models often measure responses from only one group of people by using a variety of predictive techniques to draw conclusions about their behavior. The limitation of this model is that no control group exists; decision trees and logistic regression only show the response of an audience exposed to the communication. To test persuadability, an uplift model maintains a randomized control group of test subjects not exposed to your marketing campaign.

Instead of forecasting whether a prospect will take an action once they're contacted, persuasion modeling attempts to measure whether customers will take an action only if they're contacted.

Building an uplift model means dividing prospects into four broad categories.

The first two are less interesting because their behavior is unchangeable for the purposes of a response model:

  • 1) Sure things: Prospects who will take the desired action regardless of exposure to marketing communications. In politics, these customers are inexorably partisan; in marketing, these respondents would be classified as distinctly brand loyal.

  • 2) Lost causes: Prospects whom marketers are resigned to losing regardless of the effectiveness of marketing activities. These customers feature the same unchanging mindset as sure things; they simply sit on the opposite side of the spectrum.

The two remaining groups are of much more interest to marketers:

  • 3) Sleeping dogs: These prospects could actually respond negatively to marketing communications, and change their behavior to an action that's at odds with the intent of your efforts. As the old adage goes, it's best to let sleeping dogs lie.

  • 4) Persuadables: prospects who will respond positively to your marketing communications. Naturally, identifying this segment is the highest priority of marketing departments, because the other 75 percent of the market either won't respond or doesn't need or want such contact from your organization.

  • <. l="">

    Persuadables are where marketers find their return on investment.

    Uplift in Action

    In 2012, the Obama campaign's persuadability model involved 300,000 volunteer phone calls that created a treatment group (the opposite of a control group). Out of those 300,000, the campaign was able follow-up with and survey approximately 20,000 people.

    The results of these follow up calls showed the effectiveness of the campaign's model: respondents were 4 percentage points more likely to vote for Barrack Obama after receiving a phone call from his campaign.

    Uplift modeling is often used for campaigns where consumers will be directly contacted, such as telemarketing, or direct marketing through e-mail or physical mail. Uplift models can be particularly useful in conjunction with proactive retention efforts to reduce customer churn.

    Churn models can predict the likelihood that customers with a particular set of characteristics will switch from their current service provider, and even which interactions between a customer and a service provider can cause a customer to leave. However, broad retention campaigns deployed in response to churn predictions can have negative side effects (which the sleeping dogs segment of uplift modeling accounts for). Using uplift modeling to augment the predictive analytics of a churn model can improve the precision – and therefore the effectiveness – of any retention efforts.

    In addition to retention efforts, uplift modeling is often effective in demand generation, particularly in up-selling or cross-selling situations. In the case of high-priced financial products, using uplift can dramatically reduce the amount of resources dedicated to up-selling initiatives through more precise targeting. In this vertical, purchases are few and far between, meaning most marketing is wasted on uninterested parties. Because the incremental returns are so low, reducing the volume of communication is one of the best ways to save resources.

    Uplift modeling isn't without its limitations: a baseline amount of data is required to draw significance, and deploy an effective uplift strategy. In some cases, the cost per conversions may actually rise because calculations no longer feature the sure-things segment that artificially inflates the success of marketing tactics.

    Regardless, uplift modeling has proved effective for different forms of direct communications -- from campaign cold calls to financial up-selling -- and this form of analytic modeling looks set to be a cornerstone of direct marketing campaigns across numerous verticals.

    Zach Watson writes for TechnologyAdvice about business intelligence, healthcare IT, and gamification. You can reach him on Google+ or at [email protected].

TDWI Membership

Get immediate access to training discounts, video library, research, and more.

Find the right level of Membership for you.