Why Predictive Modeling is Better Than RFM

By Lynn Dougherty, Editor In Chief, “Cowles Report on Database Marketing,” July 1998

This article is based on a speech by Jim Wheaton, a Principal at Wheaton Group. 

Catalog Age magazine's June 1998 Benchmark Survey on lists and databases shows that at least half of all catalog respondents rely mainly on RFM (recency, frequency and monetary analysis) to make marketing decisions.  But direct marketers can do a better job of targeting using statistics-based predictive models, says Jim Wheaton, vice president of strategic consulting at Customer Management Services – and he has the data to prove it.

His argument centers on the degree and depth of file segmentation.  In simplest terms, predictive modeling segments a file more deeply than RFM, to give marketers more detail about buyers' behavior.

Wheaton uses a client to illustrate his point.  He first explains how RFM was used by the marketer in the past.  With regard to recency, "The client was segmenting its database into four groups:  those who had bought in the past six months (and were the most likely to buy again); those who hadn't purchased for seven to 12 months; those who hadn't bought for 13 to 24 months; and those for whom it had been 25 months since their most recent purchase."  The client then took those four cells and divided them in three ways by frequency:  those with one, two, and three or more lifetime orders.  For the monetary part of the equation, the client used average order size, and split the file into three parts:  low-range, mid-range and high-range average order size.  All totaled, there were 36 different cells (four recency factors multiplied by three frequency multiplied again by three monetary).

For this client, the break-even mailing cost was $1.25 per piece mailed.  Also, the client generally re-mailed several weeks after the initial drop, getting a response rate that was about one-half that of the first one.  For example, if an initial drop generated about a 2% response, the re-mail would do about 1%.

To identify customers to promote to, the client first looked at past-12months mail results and determined an average response by cell, then ranked the cells from highest response to lowest.  Those cells that fell above the $1.25 break-even mark would be worth mailing; those that didn't would not.  For instance, cell number 1 consisted of customers who bought in the past six months, bought at least three times over their lifetime, and had high average order sizes.  Those people would generally bring in $4.00 per piece mailed.  They would also get mailed twice, since they were likely to generate $2.00 per book from a second mailing, which was still well above the $1.25 break-even mark.

"But take a look at a customer who hadn't bought for 13 to 24 months, had only bought twice and the average order size was mid-range," Wheaton says.  "Customers in that group generated $2.00 per book, but could only be mailed once, because a second mailing at 50% response would only generate $1.00."

Far, But Not Far Enough
But his client's RFM analysis didn't go deep enough to identify those customers with the most potential, Wheaton says.  One reason is an insufficient number of categories.  "Customers who purchase 10 times, for instance, were categorized the same as those who purchased three times, when in fact they were much more loyal and should have been noted as such," he explains.  "If we increase the categories of RFM, though, you don't have a large enough sample size on any given cell to get a realistic read.  That's your basic problem with RFM."

Statistics-based predictive models, he says, don't have these limitations, and "also do a better job of finding the extreme ends of the scale – the really good and bad customers in terms of future behavior."  For this same client, Wheaton's organization built a predictive model to supersede the RFM cells.  With predictive modeling, he explains, "you take into account additional information that has value, such as demographics and customer satisfaction indicators, like the number of product returns or back-order complaints.  You then add or subtract points for each one, and total the points for each customer."

Everyone on the database is scored with the same point scale, sorting from high to low.  The entire database is separated into 10 equal deciles, with decile 1 as the best customer group.

According to the figures in the chart below, customers in decile 1 generate roughly $8.14 per piece mailed, while decile 10 only brings in $.44.  Combined, all deciles do an average of $2.49 per piece mailed.  The lift column shows the performance for each decile compared with the $2.49 average, while the last column shows a cumulative running lift.  For instance, if you mail all segments above the $1.25 break-even mark – segments 1 through 7 – you're indexing at 130, which means you would do 30% better than the $2.49 per piece revenue average.

Customer Predictive Model:  Depth of File

After Modeling…What Next?
Even after modeling, there are still nuggets of golden opportunity buried within a database, Wheaton says, that can be uncovered through tree analysis, which highlights the variables that separate people who are likely to buy vs. not likely.

Suppose that the majority of your audience is women, and that when your file is split male-female, the women spend an average of $2.88 per book, while the men spend only $1.42.  In the case of the client above, Wheaton notes, tree analysis on all males in the file showed that men who buy jewelry bring in $2.95 per book, 'while men who buy all other merchandise only do $.89.  Without tree analysis, this client might have continued to focus only on women.  But with this information in hand, the client can now create targeted programs to men that cater to their special needs.

Jim Wheaton is a Principal at Wheaton Group, and can be reached at 919-969-8859 or jim.wheaton@wheatongroup.com.   The firm specializes in direct marketing consulting and data mining, data quality assessment and assurance, and the delivery of cost-effective data warehouses and marts.   Jim is also a Co-Founder of Data University www.datauniversity.org

Top >>

Copyright © 2004 Wheaton Group LLC. All rights reserved.