Advertising Week


Advertising Week 10th Anniversary Official Guide

Issue link:

Contents of this Issue


Page 239 of 333

238 BRAND ED The intersection that crosses this street is deterministic v. probabilistic. Most first party datasets are assumed to be deter- ministic—because when a call center receives a customer's call, we assume the customer is giving their correct identity. However, there is a lot less certainty about third party data, which sometimes is as- sumed to be deterministic but often is a mixture of deterministic and probabilistic, the latter being modeled. Modeling data requires a truth or training set of high quality data which has a set of generally available linking vari- ables, often demographics, and unique media, consumer or behavioral data such as purchase intent. For example, demo- graphic predictors of purchase intent are derived from the training set which has known purchase intent. Those predictors are then used to model purchase intent in other, often much larger databases such as the cookie pools that may reside on a Data Management Platform. Which is better? Perhaps if you knew your deterministic data was really deter- ministic you might conclude it is better. However, often you don't, hence sev- eral industry associations are exploring guidelines and standards for looking at the true accuracy of deterministic data sold by third parties. There are likely two dimensions that will ensure that databases of the future will be hybrids of deterministic and prob- abilistic data: the eventual rise of Multi Touch Attribution (MTA) and Privacy. According to a CMO/IBM survey reported in December, "better connecting campaigns into a comprehensive, connected experi- ence", the Omni-channel experience was the highest priority for marketers today. MTA is often contrasted with Market Mix Modeling(MMM). Critics of Market Mix argue that the largely regression based ap- proaches have not kept up with advanced techniques of machine learning and artificial intelligence. We hear on the street, "What's the value of shelf space when store traffic becomes e-traffic?". To be fair, modelers such as Nielsen have recently announced breaks with tradition with the introduction of machine learning and some AI like features in their Market Mix analysis. Market Mix is top down—predicting sales from aggregate ratings, pricing, competition ad spend and many other marketing variables. However, one of the fastest growing sectors of MarTech is addressable tar- geting to individuals. One study reported in ADWEEK, Nov. 17, 2016 suggests ad- dressable TV will double by 2018. Even linear television will join the wave as IPTV enables it. MTA will be the go-to database for Omni-touch campaigns and MTA will require a mixture of deterministic and probabilistic data. MTA databases aspire to be a matrix of media, consumer, demographic and behavioral data for every household in the country. The problem is that these matrices are sparsely filled with deter- ministic data. A few million return path households that can be directly matched to a database of a hundred million loyalty card households leave you to probabilis- tically model the remaining 95%. Hard data become the training set to model the remaining loyalty card households. Together, direct matches and the models they permit allow marketers to more pre- cisely target the whole population. The second dimension likely to drive the hybrid database is privacy. Most companies with significant first party data realize that it has monetary value. The mistake that they often make is that they believe it only has value within their own industry. As cars become more con- nected, the driving data that auto makers collect has value to industries as far afield as municipal governments and quick service restaurants. The problem we have seen is that in many B2B agreements and many B2C subscriptions, deterministic or direct matching of individual names for the purpose of addressable targeting of in- dividuals violates contracts, subscriber agreements or corporate privacy poli- cies. At this moment, companies facing these restrictions may allow DMPs to use their first party data to develop segmen- tation algorithms such as "quick service frequenter" under the condition that ev- eryone in the DMP data platform must be assigned the probabilistically modeled behavior, even where the true determin- istic behavior is known. In this way, the first party data is used as a training set but never used to target an individual. It is some time before DMP's and MTA data platforms have hard data Omni-touch data on even a quarter of the population. Until that time and as modeling permits addressable targeting without retention of privacy sensitive data, it is certain that a direct data match is the best way to train algorithms and probabilistic models are the best way to extend that information to the broad population in the most privacy conser vative way. • I was talking with Dennis Buchheim, SVP at the Internet Advertising Bureau a few weeks back. He told me the two issues that he most often hears from members are: "identity and data quality". Specifically, members are asking for guidelines and standards for consumer identity and audience profile quality.

Articles in this issue

view archives of Advertising Week - AWNewYork_OfficialGuide-2017