Betting the Farm on RFM, Part 1

It wasn’t a subject I knew much about, at least not until a recent conversation with my friend Jim Novo, former vice president of marketing and programming at the Home Shopping Network and co-author of “The Guide to Web Analytics.” When Jim wants to discuss anything relating to metrics and modeling, I listen closely.

Jim had phoned because he was concerned that some articles written recently about RFM, including pieces by my fellow ClickZ columnist Mark Sakalosky, were a bit off the mark.

RFM Served Up Right

“What in the world is RFM?” I thought. The abbreviation itself turned out to be pretty straightforward: recency, frequency, and monetary value.

Recency represents the number of days since the customer last completed the action you are profiling. Frequency represents the number of times the customer has completed this action since the first time she completed it. Monetary value represents the total value (usually total sales) the customer created by completing these actions.

The classic RFM model produces “scores” that rank customers relative to each other for the likelihood that they will repeat whatever action is being profiled. Any action can be profiled — visits, purchases, log-ins, and so on. High likelihood to repeat an action, providing this action has economic value to the company, means high future value. Low likelihood to repeat means low future value. RFM is that simple.

What’s the Buzz About?

Using RFM, you can predict the future value of customers and their likelihood to respond to promotions. Jim piped, “The average marketer should be drooling over this. If you can predict future value and likelihood to respond, you have the magic wand to design high ROI campaigns.”

Jim’s right. Two years ago, he published “Drilling Down: Turning Customer Data into Profits with a Spreadsheet,” at a time when nobody wanted to hear about marketing to existing customers. Customer acquisition was all the rage. “This same focus on acquisition-while-ignoring-retention happened with cable TV, cellular phones, and TV shopping — ultimately leading to high customer churn rates and business failure,” said Jim. “RFM proved to work even better online than it did offline. But at the time, nobody cared.”

The Model, Please

Here’s the way it works. First, you figure out recency (typically in terms of number of days since last action), and sort all customers on this basis, with the shortest number of days since last action at the top and longest number of days since last action at the bottom. This list is then broken into five groups (called “quintiles,” if you really want to know). The top 20 percent you label “5,” the next 20 percent down the list you label “4,” next 20 percent “3,” and so on. Customers labeled 5 are the 20 percent who completed the action most recently; customers labeled “1” are the 20 percent who last completed the action the longest time ago.

Then perform the same sort of organization with frequency. Sort customers by most frequent performance of the given action to least frequent, and label the top 20 percent 5, the next 20 percent 4, the next 20 percent 3, and so on.

Then do the same thing for monetary (total sales), if you have this figure (more on this later).

At the end of this process, each customer then has a three-digit RFM score, ranging from 555 for best customers (top 20 percent in recency, frequency, and total sales) to 111 (lowest 20 percent in recency, frequency, and total sales) and everything in between, for a total of 125 possible states (5 x 5 x 5 = 125). This gives you a good amount of workable detail without being overwhelming.

RFM allows you to segment customers by their future value and likelihood to respond. If you have 125,000 customers, each cell or ranking would have 1,000 customers in it. It’s a beautiful thing for marketing purposes, don’t you think? You have automatically created visitor segments of equal size, the perfect setup for direct and database marketing test programs.

If you don’t have an M for a behavior — such as visiting a Web site and ranking the likelihood of a return visit — don’t worry. You can use RF scores, in which case you will have rankings from 55 down to 11 (5 x 5 = 25). M is the least powerful variable anyway; 80 percent of the predictive power of RFM is in RF by itself.

Will the Real RFM Please Stand Up?

Now here’s where Mark and I differ.

I believe using three segments for each variable (R, F, and M) “dumbs down” the model to an unacceptable level. Doing so give you 27 (3 x 3 x 3 = 27) customer segments instead of 125. Will it work? Sure, but it’s suboptimal — unless you only have a couple thousand visitors or buyers. What’s the point of a ranking system that loses accuracy the more data is added to it? I’m not interested in dividing my world into just small, medium, and large — I want to identify all the profitable segments, wherever they may be. By collapsing the detail in this way, you lose predictive power in the model.

Some people use average purchase amount for M instead of total sales. What if the customer is a very long-term, loyal customer and, for whatever reason, buys low- to midpriced stuff? These people can be very responsive and have high future value. Yet by using average purchase price for M in RFM, you would snub these loyal long-term customers who, in fact, have very high total sales.

By using average purchase amount, you are demoting high-value, low- or midticket buyers to a level relative to that of one-time or infrequent buyers who bought high-ticket items. This undermines the predictive capabilities of the RFM model.

Some people try to defeat the relative ranking system of RFM, but the beauty of the RFM model is the relative ranking system. You need to let the data decide how the segments should be divided up rather than imposing some arbitrary order on it yourself (a point made by Mark in the second part of his series).

For example, if you group a bunch of customers into a “made a purchase more than three months ago but less than nine months ago” bucket, you are not allowing the ranking system in RFM to do the job it was designed to do. You’re forcing all of these customers to have the same score, and you are bound to create segments of different sizes with widely disparate behaviors. You lose the natural ranking order of the model. Who is to say the cutoff should be at three months? You want to know how customers compare with regard to future value and likelihood to respond so you can target campaigns to the highest potential value customers. Let the data itself determine where the break should be, and let RFM take care of itself, please.

So, just what can you do with RF(M)? How does it work in practice? We’ll tackle that next week. I hope you’ve visited the other sections of ClickZ recently and that you return to this column frequently. Unfortunately, there is no money involved.

P.S. For those of you in the Denver; Newark, NJ; Detroit; and Atlanta areas, I’ll be doing full-day workshops on conversion rate marketing. I hope to see you there.

Related reading

metrics hp
Report: millennials are more likely to share an ad (but also to mute it)
A graph showing a small pile of money at one end and a larger pile at the other, with an ascending line connecting the two.