UPDATING THE NEW DM: Working With Tricky Segments

IN A RECENT COLUMN, I suggested that modelers could improve their results by splitting data sets according to some critically important variable such as tenure (the length of time a customer has been on the file), then building separate models for each major segment (“Why One Model Is Probably Not Enough,” DIRECT, December 2000). The reasoning is, it’s intuitive that the usual set of modeling suspects (recency, frequency, monetary value, products purchased, source and the whole set of demographic variables) will display different relationships with response or sales, depending on the tenure segment. Adding tenure as a variable without taking interaction into account isn’t sufficient to capture its full effect.

As if this isn’t complicated enough, I came across an article in the October 2000 Journal of Marketing titled “On the Profitability of Long-Life Customers in a Noncontractual Setting: An Empirical Investigation and Implications for Marketing” by Werner J. Reinartz and V. Kumar. The authors question what they consider fundamental direct marketing beliefs. This includes the opinion that there’s a strong positive relationship between customer lifetime and profitability in a noncontractual relationship.

In other words, Reinartz and Kumar think direct marketers and other academics believe customers who hang around a long time, buying every once in a while, are profitable. So, they say, every effort should be made to enhance the relationship between buyer and seller.

Of course, DMers who’ve looked closely at the data know that the costs of servicing infrequent buyers may indeed exceed the margins they yield. And the authors discovered for themselves that the simple relationship between lifetime months on file and lifetime profits is relatively weak (r = about.2 for the two groups studied).

Reinartz and Kumar studied some 9,000 households over a three-year period. They split them into two cohort groups – January and February 1995 starters. What I found interesting and potentially actionable was the discovery that the authors could divide a significant number of these catalog customers into four major categories:

Segment 1. Those with relatively long active lives and high lifetime revenue.

Segment 2. Those with relatively long active lives and low lifetime revenue.

Segment 3. Those with relatively short active lives and high lifetime revenue.

Segment 4. Those with relatively short active lives and low lifetime revenue.

The graph below indicates that customers in segments 1 and 3 look somewhat alike and behave in a similar fashion over the first 12 months, then begin to separate over time. No doubt this is true. The question, however, is: Can this disparity be predicted, and predicted early enough in the buyer’s lifetime that corrective action can be taken?

The authors argue – correctly, I believe – that simple RFM analyses will miss this phenomenon. Also, catalogers, as a consequence of not understanding that their database consists of these segments, will overspend on the short life/high revenue segment before traditional RFM analysis will depress mailings to it.

So the key question for catalog marketers is: If this effect is widespread – if there really are customers who come in for a short while, buy a lot of merchandise and then leave – can they be detected? Will modeling tenure segments capture this effect? Probably not, at least not by itself.

What might work is a principal component analysis of the available purchase behavior data over the last six months. This approach might discern either a trend in dollars spent, or a trend in the particular products purchased. This would indicate that the customer was displaying a pattern associated with customers who buy heavily for a time and then switch to someone else – for reasons we can only speculate about. The authors’ own attempts to predict buyer behavior were only modestly successful. But as they say in the academic world, more research is needed. And they’re probably right.

Source: Journal of Marketing, October 2000. (Values are approximate)