Exceptional results.
More on Fuzzy Segments
By: David Shepard
This article first appeared in Direct Magazine
In our last article in this space we argued that segmentations based on a simultaneous combination of survey generated demographic, behavior and attitudinal data tend to produce a small number of interesting, fun to name but often fuzzy segments. Fuzzy in the sense that while the segments differ with respect to the average value of important segment defining variables, the spread around the averages can be significant.
For example, Segment 1 may include older married couples, who are frequent buyers, and who prefer to shop at retail; as opposed to members of an adjacent segment, Segment 2, who are also, on average, older married couples, also frequent shoppers, but who prefer to shop by mail.
Now, the problem is that included in Segment 1 will be some older singles and some younger married couples, who got into that segment because of their shopping patterns and their channel preferences closely matched those of members of segment 1. This result is what we mean by a fuzzy segment.
So, what's the problem with fuzzy segments?
The problem with fuzzy segments is that very targeted promotions -- promotions that implement creative executions that act on the assumption that all of the members of a segment are identical to the description of the average segment member -- frequently fail to beat control packages that have been fine tuned to appeal to a broader audience. Or, when the response rates to targeted promotions do beat the response to a control promotion, the increase in response is oftentimes not sufficient to justify the increase in costs associated with the entire segmentation effort.
This observation that "our segmentation strategy didn't beat our control strategy" is frequently voiced at direct marketing conferences or seminars on segmentation - not by the speakers, of course, but by the practitioners in the audience who have gone down this path and failed.
So, what's the solution?
The statisticians might argue that more segments are required to solve the fuzziness problem. And, that's true as far as it goes. The programs that produce segments will let you go pretty far before they run out of potential splits. But then you run into another difficulty, and that's building a model that can accurately predict segment membership from the information contained on your database, information you have about everyone on the database, not just information taken from the questionnaire that produced the segmentation.
Remember that you generally can't afford to administer a multidimensional questionnaire to your entire database of customers or prospects. And, the problem is that the greater the number of segments produced by the segmentation process the harder it is to model segment membership based on available database or purchased overlay data. In a future column we'll discuss the success we've been having using CHAID to place customers into segments when the traditional method, discriminant analysis, has difficulty predicting segment membership.
For these reasons we favor an approach that attempts to keep things simple, at least initially, by isolating the three dimensions of segmentation (Behavior, Attitude & Demographics) creating three separate segmentations, and then using the right combination of segments, in combination with predictive models to drive contact strategy.
Of course, many direct marketers have relied solely on predictive models to drive their promotion decisions, and have sometimes used profiles of the best and worst performing deciles to help shape their creative decisions, but this 80's strategy won't cut it in the 90's.
After all, regardless of what business you are in, it's hard to argue that there are no alternative creative strategies for the different demographic or lifestage groups that comprise your customer file; or that from a creative perspective that your predictive model captures all of the relevant behavior dimension; or that given the results of your attitudinal segmentation, assuming that you have done one or were to do one, that you would still see only one creative strategy worth pursuing.
In fact, given the range of products and creative executions available to you -- in your mailing pieces, or over the phone -- it's more than intuitive that knowing your customer's demographics, and/or their prior purchase behavior and/or their predisposition to buying from you or staying with you, given a choice, would make all the difference in the world.
Just think about: a cable system selling Internet access; a bank selling securities; a department store selling travel and related services; a cataloger trying to reactivate a dormant customer; or a book club deciding which titles to offer in the next catalog. In these situations knowing as much about the customer as possible is essential. And, this is especially true in situations where you yourself are not generating a promotion, but instead are preparing to respond to a customer initiated contact, at an ATM, at a customer service center, or at your web site.
So, how does one go about implementing a manageable strategy that employs three separate segmentation schemes and a host of predictive models, as well as ad hoc selection capabilities.
A starting point would be do identify a set of perhaps 10 to 20 intuitively important demographic variables and build a demographic segmentation resulting in between four to six clusters. Similarly one could identify a similar number of key behavior variables and create anther four to six behavior based clusters. The next step would be to reach agreement on the attitudinal parameters you want to measure including, almost certainly, measures of loyalty, share of requirements, customer satisfaction and brand equity. Then field a survey and develop a set of attitudinal clusters and through modeling assign all of your customers to the most appropriate segment.
At this point you will want to size each of the resulting combinations and combine them or roll them up into a manageable number of three-dimensional marketing clusters. The decision regarding which clusters to combine will have to do with size and marketing strategy. Obviously there's no point in creating clusters that that represent less than one percent of the database, but you may start off by identifying 50 to 100 clusters and then roll these cluster up into a dozen or so major marketing clusters (in much the same way the major commercial clustering products build their segmentation products).
A major marketing cluster should make sense in terms of marketing strategy. Think about it this way: with regard to major promotion initiatives, if you can't think of a marketing strategy for one cluster that would be clearly different from the strategy for another cluster, then for all practical purposes you could combine the clusters.
So, here we are with say a dozen major three-dimensional marketing clusters. What do we do now? First of all we don't act as if we only have a dozen clusters and therefore have the same strategy for everyone in the same cluster. For each cluster we need to build models to predict different kinds of behavior: attrition, response, lifetime value, etc., at the individual customer or household level so that we are in a position to decide how deeply to promote within a cluster or how frequently to promote to customers within a cluster.
Additionally, for some selection decisions we may require the ability to execute rules that require access to both major and minor cluster assignments; individual one-dimensional segment assignments; individual variable values within a single dimension and model scores. In other words your database should house the results of all the intermediate segmentations that were required to arrive at the final marketing cluster assignment.