Essentially, user scoring and propensity modelling are the same thing and they also go by a number of other subtly different synonyms, just to make everything more confusing.
You might have heard or seen reference variously to:
- probability scoring
- best-next-action modelling
- continuous scoring algorithms
- algorithmic attribution
- lookalike modelling
- predictive analytics
- customer behaviour scoring
- propensity score matching
However you refer to it, propensity modelling is changing dramatically as machine learning is lending its weight to improving the efficiency of advertising and marketing.
Indeed, Chris O’Hara from Krux refers to this revolution in data science as the most important trend in programmatic advertising (see Econsultancy’s programmatic trends for 2017). Here are a few of Chris’s most succinct comments on the issue:
- “Getting a programmatic edge means…being better than your competitors at knowing where and how much to bid…”
- “…most marketers and agencies have little native competence in user scoring and propensity modeling…”
- “[But] we are starting to see…platforms that embed machine learning and artificial intelligence into their user interfaces in such a way that business users can access such capabilities without…having statistical abilities.”
So, what is it?
Advertisers want to use the profusion of online data to their advantage, creating a model of certain consumers in an effort to predict who will respond to specific messages, or indeed buy a particular product.
Data science now allows for meaning to be found in enormous datasets, which is essentially what machine learning does, a form of AI that can ‘learn’ from each new interaction between consumer and message.
Propensity modelling chiefly refers to the modelling of a person’s propensity to click on an ad or to convert (once they have clicked). Whilst this type of modelling has historically been applied to a company’s own customer base in order to predict who is likely to churn for example, or which sales lead may bear fruit, it is now being used on a wider scale online to target unknown consumers, too.
This technique helps to plan ad impressions based on a whole range of variables and it can now be conducted in real-time during the bidding process for online media (through integration of data management platforms and demand side platforms).
Advertisers can target different behaviours and demographics, or create segments based on various personas of their current customers (high value buyers, discount buyers, browsers etc.) and these segments can be used in lookalike modelling to define what traits to look for in a larger data set of visitors.
Understandably, those companies with rich first-party data are better placed when it comes to accurately modelling potential customers, because they can describe a more accurate portrait of their current customers.
The increased ability to track customers across multiple online channels and, via a device graph, from desktop to mobile, means that propensity modelling can be brought to bear on more than just display advertising.
Scoring across more than just display
In our 2017 programmatic trends roundup, Chris O’Hara picks up on the complexity of scoring across multiple channels:
“[The use of machine-learning powered platforms] was more straightforward when it was just available to display marketers seeking to manage bid pricing thresholds on cookies.
“[However,] today, marketers are increasingly using data management technology to map users across their device graph, and expect the ability to score users against their interactions across every addressable channel—not just ‘display’ advertising, but also email, commerce, and website experiences.
“To do this correctly, marketers need to map users to all their devices and be able to store highly granular attribute data going back longer than the life of the typical cookie.
“These are “big data” problems that require highly advanced technology. Much of what is happening today is ad hoc reporting in spreadsheets that drives manual optimizations across many different buying platforms.”
In the long run of course, marketers are trying to map the customer journey to purchase, and to tickle as many consumers along it as possible. For example, marketers want to understand at which point in the customer journey particular content versions and formats have the most positive impact on the decision to purchase.
However, propensity modelling and increased targeting doesn’t, of course, mean 100% efficiency. Quality of ad placement is vitally important, and private marketplaces allow advertisers to take part in invite-only auctions for the best ad spots that will give the best return.