The goal of any marketing strategy is to reach the right customers with the right message at the right place and time. In order to accomplish this, marketers need both the right information to make informed investment decisions and the ability to serve messages to specific audiences with as little waste as possible. Until recently, media investment was an “all or nothing” proposition. Marketers had to buy advertising adjacent to content based upon the known information about “who” was in the audience. Demographic measures were the best common currency we had to make these choices. And, once an investment was made, the marketer’s message was placed adjacent to the purchased content and everyone in the audience received the same message. We knew there was a lot of waste with this approach, but it was the best we could do.
Advancements in programmatic digital media planning and buying changed all this. Now, marketers no longer have to buy the total audience associated with a particular piece of digital content. In programmatic digital display, marketers can now buy only the specific “Cookies” that show up on a publisher’s site that fit their selection criteria while leaving the remaining audience open to other bidders who may find them a better fit with their own interests. You would think this improved ability to target only select Cookies with specific messages would radically improve the return on investment of digital display advertising. Unfortunately, the results have been mixed at best. Despite using all forms of Big Data to get a more precise understanding of the Cookie and the recent online behavior associated with it, click-through rates and digital display return-on-investment still remain relatively low.
Addressable television advertising is on the horizon and it will begin to level the playing field: marketers will be able to target advertising messages down to the individual household level. As the technology advances, marketers are left with an important question: Which data are most important for making informed household targeting decisions? In the old model, we didn’t have much choice. We had to use the demographics of the total viewing audience for TV networks and programs. Now, we have all kinds of Big Data at our disposal to help inform our decisions. Perhaps the default approach will be to take what we are currently doing in digital display programmatic and apply it to programmatic TV advertising. However, the lessons learned from digital display programmatic to-date suggest that the current approach leaves room for improvement. There’s a lot we still don’t know.
The research team at Medill’s Spiegel Research Center (SRC) acquired data sets that detail voting behavior in the recent Democratic and Republican presidential primaries in the state of Texas. We were interested in exploring the factors that would predict whether a household would vote in the Republican primary (Red) or the Democratic primary (Blue). To answer these questions, we acquired three different data sets: set-top-box cable TV viewing data, primary voting data (Red vs. Blue) for the same universe, and detailed household demographic data. Voting is a key outcome in political advertising, and voting for a party can be viewed as purchasing its product.
Our first step was to determine whether more detailed household demographic data would help predict Red vs. Blue voting behavior. Using a rich set of demos in machine learning models, we can predict party affiliation with AUC = 0.66, where AUC is a measure of predictive accuracy and the value 0.5 is random guessing. This outcome suggests that demographics do, indeed, matter. And, the smarter we can get about the demographic characteristics of individual households, the more successful our addressable efforts will become. One could say, “The more things change, the more they stay the same. Demographics still matter.” [FIGURE 1 – The AUC Curve]
Does the story end with a better understanding of household demographics? Or, is there something about the television content, itself, that can lend valuable clues? What if what you view is as important as who you are in determining your voting behavior? To find out, the SRC researchers created approximately 4,000 variables measuring viewing behavior and included them in the same machine learning models. The AUC measure on a test set that was not used in training the model improved to 0.88, which is an impressive improvement. Sure enough, Red-voting households have many unique viewing preferences in comparison to their counterparts in Blue-voting households. Here are a few highlights… [FIGURE 2]
The implications are quite significant. If marketers focus solely on the demographics of households, they will miss the potential to deliver a much greater return on their investment from television. Our findings suggest that the combination of rich demographics and rich viewing data are the perfect combination for success in this new addressable world.
Written by Ed Malthouse and Judy Franks, Northwestern Medill IMC Faculty
Edited by Yunqui Zhang, Medill IMC Class Of 2017