Blog post

I Predict: The Rise of Predictive Analytics

By Martin Kihn | March 11, 2014 | 2 Comments


You don’t have to be Nostradamus to predict that predictive analytics is going to become more and more important to digital marketers. Gartner sees an ongoing shift from analyzing historical descriptive data in aggregate to tell the story of “what happened,” to performing calculations on data sets to predict — with more or less confidence — how a particular individual will react to a particular message, offer, or recommendation. That insight can be extended, using different math, to find people similar to the people who are likely to respond . . . until, at some point, all of us are presented only with marketing communications to which we are likely to respond.

At which point we will be powerless to resist the marketer. Or so they hope.

Predictive Analyst
Predictive Analyst

The predictive analytics surge is being driven by two forces: (1) the higher velocity of data processing, driven by frameworks like Hadoop, and (2) the increased availability of data at a customer level. By “customer level,” I mean the emergence of individual profiles — which can be either personally identifiable (in the case of a known customer), or anonymous yet unique (in the case of a site visitor or individual encountered on the web).

Predictive analytics as a discipline is not new. It has decades of applications in the fields of business intelligence, for example, predicting how likely a person is to default on a loan, or credit card. Actuarial science in the insurance industry is an example of predictive analytics with its own university degrees. However, its use in digital marketing has been more recent and narrowly applied. This situation is about to change.

To take a simple example, I logged into LinkedIn recently to see who was trolling my profile and was confronted with the following example of predictive analytics in action:


Notice (1) how specific these recommendations are, and (2) how easy they are to act upon. (“Follow Ryan Holmes and get +1% more page views!”) More than best practices, these recommendations come with a predicted impact variable. Also notice the class of persons the recommendation applies to (“Marketing and Advertising professionals like you”), which implies at least some of the recommendations hold true for anyone in that group. What we are seeing here looks like the output of a regression model familiar to any MBA, with the independent variable — the one we’re trying to increase — being page views.

Who cares? My point is that LinkedIn has had this data all along, but is only now deciding to offer it up to drive action. There may be another motive, as my next visit to LinkedIn served up a different set of recommendations:


One of the recommendations, to follow an agency, has no impact associated with it, but presumably helps LinkedIn with its own engagement efforts.

A similar feature is being introduced by social analytics vendors, who are providing predicted responses to social marketing posts before they are posted, based on factors like format and message length. A typical automated feedback message might be: “Add a picture and shorten the text to 12 words to optimize engagement.”

Both LinkedIn and the social analytics vendors are performing predictive calculations — including regression, classification,  decision tree, and clustering models, — and providing the output to users. The result is a short list of specific, data-driven actions that are likely (though not guaranteed) to improve your efforts.

I predict all digital marketers will have to say to this development is: Like button.

Comments are closed


  • From the fact that it took LinkedIn so long to surface some of these recommendations you can see that predictive analytics can be difficult to pull of (especially if you are doing it all in house). However, ironically predictive analytics for many marketers can also be easier than the traditional way to segmenting and targeting prospects and customers. Traditionally, segmentation involved many rules that manually need to be configure and updated to come to groups of people and personas that have an affinity for the same things. For example, in the old days, if you wanted to try and figure out who was likely to buy you might setup rules that would look at website behavior and email opens or recency, frequency and monetary value of purchases made (RFM). With predictive, instead of thousands of manually configured rules, all you need is one rule: “who is most likely to buy” and machine learning will do the rest. That’s much simpler. I mention this because I have come to realize that for marketers, predictive analytics shouldn’t be the “next big thing” but the “first thing” that is not only a lot more effective, but also a lot easier that the segmentation available until now.

  • Martin Kihn says:

    Thanks Dominique. That’s a great example of the real value software can bring to the calculus of marketing — it can remove a layer of “best guesses” (which are often used to construct the rules) and manual processes with an automated, iterative system that frees the marketer to do what they should be doing, i.e., communicating with people.