We Know Predictive Modeling is Useful for Predicting Future Behaviors, Can You Give Some Examples of how a Model Works?
Jeff:“Predictive modeling looks at a variety of data elements and builds an algorithm off of all those different elements and scores it against that universe. For example, a lot of times in retail someone would say our target audience is aged 45-55 with a household income of $75k, living in the suburbs or something like that. Those are hard cut selections. You have to meet those 4 things. You’re either in or out. When you build a predictive model you might end up with something that kind of looks that way but you would look at a variety of data elements. It could be 300 elements and you come down to 30 and end up including someone that might be 28, not live in the suburbs, but has other attributes because of the data elements – they opened email or they clicked on something or they recently bought at a competitor or whatever that data might be that throws them into the model. Now they’re somebody you want to target. So the beauty of an algorithm is that looks at all these elements and provides a score off of all the elements versus just having very hard cut rules on say 3 or 4 or 5 or 6 hard selections.”
Sophia: “The model also ranks the importance of those factors. Say female versus male or age or income. Which one of those are most important to you? Maybe the gender is most important and then age comes next and what’s the ranking of these and what’s the optimal weight we should put on each of these factors. So basically the model arrives at the optimal weight among all these factors.”
Jeff: “With the model you would forecast and predict what you think those groups or individuals or segments how they would perform. So you might say segment 1 would get a 2% open rate let’s say and then we would actually track that against it and go ‘wow you got a 9.875 open rate so it was pretty accurate. Or this group here we thought was going to be not so hot so maybe try something different and open rate is only going to be 1% and you track that group and the accuracy of the model and its ongoing performance and you kind of refresh it and make sure the model’s still performing the way you think it’s going to.”