Understanding the Basics of Predictive Modeling

Understanding the Basics of Predictive Modeling

By Don Bottelson

In marketing, terms like “Modeling,” “Predictive Analytics” and “Profiling and Segmentation” get thrown around a lot, and sometimes these terms aren’t self-explanatory. Since this is what I live and breathe every day, I’ll try and explain some real examples of statistical modeling and how they apply to your business.

Ship builders for centuries would build complex models of sailing vessels so they could see how the rigging for the various masts and sails would or would not work together. “Modeling” is a term for building a vehicle that replicates the actions of the real world in a controlled environment.  In many cases, it is to predict how the real world will act or react to certain circumstances, or to specific stimuli. 

In marketing, what we’re trying to predict, generally, is a response to a direct marketing tactic, not the raising of a sail.  This is done by building statistical models as opposed to physical ones, but let’s not get ahead of ourselves.

Imagine you’re playing a simplified game of Craps.  You throw two dice once, if the total is 7 or 11 you win, otherwise you lose.  There are 36 total combinations possible and 8 of those are either a seven or eleven.  That equates to a 22.2% chance of winning.  But we want to better our chance at winning and have been told that there are other sets of dice that will more consistently throw a seven or eleven.  We put aside our morals for this example, ignore the devious ways someone might accomplish this task, and test out three other pairs of dice. 

We throw each pair 100 times and record the number of sevens and elevens.  In this example, we would refer to each new pair of dice as a different model and we would call throwing them 100 times “validating” the model.  The first set results in 22 seven’s or eleven’s, the second set 24, and the third set 25.  The difference between what we would get with the standard set of dice and what we’re getting with the new dice is called the “lift”, and the best set of dice gives us a lift of 2.8% (25%-22.2%).  So, we pick up that set of dice and head into the game expecting that in the real world we’re going to see a 2.8% increase in our winnings.

Marketing isn’t gambling, but it’s not hard to imagine that every piece of mail is a throw of the dice and we want to use the set of dice that results in the best response.  In direct marketing, many things go into a response, typically referred to as the list, the offer, and the creative.  While our dice example could refer to any or all three of these we’re going to focus on the “list”, assuming the other two are fixed. This is because the list is the portion of a campaign’s strategy a statistical model can have the greatest effect on. 

When constructing a direct marketing tactic, you and your customer work through audience criteria (the list), the offer and the creative.  None of this needs to change in order to use a model.  Even if your audience selection criteria are as good as they can be, unless you’re planning to mail to everyone meeting that criteria, a model can help you order the resulting universe and increase your response rate.

__________________________________________________________________________

Don Bottelson is Marketing Analytics Team Leader at SourceLink