Q: As you can see, the response rates decline fairly uniformly from 555 to 511, then rise for 455 and decline uniformly to 411, rise again for 355, and so on. That's what I implied by a sawtooth pattern. Is the above pattern fairly normal? Based on this data, can you tell if RFM is the right sequence of variables for us? How do you tell if only 2 variables (RF or FM, e.g.) are predictive and throwing in the 3rd variable messes things up?
A: Well, first let me say, if you are testing a sub-group of the population - loyalty card members - then you may see these kinds of shifts. Loyalty card holders especially are a unique group, often driven by Frequency as a result of incentives given. Hard to tell without looking at the program, but it's possible the program is modifying expected behavior. After all, that's what you do a properly constructed loyalty program for - to modify behavior! So in your data you might be seeing the footprints of your program at work.
I wouldn't say the data is "messed up", it just follows a slightly different pattern, caused either by the population selection, the loyalty program, or some combination. As long as this pattern is consistent though, you can still identify and target profitable segments.
If you would rather see a consistent "parabolic" type response curve, try sorting / scoring the Frequency quintiles *within* each Recency quintile, rather than across the entire population (if you did it this way). This will effectively boost the importance of Frequency in the score and have a smoothing effect. If you really want to go all out, then also sort Monetary *within* each Frequency quintile rather than across the entire population. This "sorting within quintiles" approach is a ton of work manually but is in effect the way professional RFM software does it.
The bottom line though is in the ability of RFM to segment customers into groups which behave predictably so that you can maximize profit. The numerical score is not really an issue; it's how people with a certain score behave. But if you have a preference as to the way the scores fall, fine.
Q: Besides these questions, where we really got tripped up was in looking at the incremental lift per customer (sales per targeted customer minus sales per control customer; the test and control group were well matched and went through statistical rigor to ensure that there weren't inherent differences in them). Here is what we saw:
Q: As you can see, the incremental lift is all over the place. Some of
the lower RFM cells show up higher in lift while some of the higher
ranking RFM cells (e.g., 554, 542, 532, etc.) show negative
lift! While I am showing the numbers from one mailing, the
results are fairly consistent in the other mailing in that some of the
mid- and low-ranking RFM cells show the highest incremental lift
whereas many of the high-ranking RFM cells show negative lift.
How do you interpret this data? What do we change? Any
advice you can provide would be very valuable.
The short answer is that RFM is a response model, not a profit model (though it can be used to construct a profit model, as you have seen). What you are seeing is normal; response does not always equal profit.
You have just proven this idea with your data but it continues to elude most marketers. The problem with sending promotions to people who are already highly likely to respond (have high RFM scores) is many would have *bought anyway without the promotion*. And the profits from low scores? Direct evidence your promotions are retaining customers and making money.
Did you read Chapter 29, Expense and Revenue You Might Not be Capturing: Subsidy Costs and Halo Effects? It's late in the book and I have found over the years many people get excited about the implementation and perhaps skip some of the later chapters...
Anyway, your "higher ranking RFM cells show negative lift" is proof positive that subsidy costs exist, and are why it often is a bad idea to do some kinds of promotions to best customers. You literally decrease their value with each promotion, because the subsidy loss is money they *would have spent anyway* without the promotion. It is quite common with these high-scoring groups, for example, to see a loss of about $4 with a $5 discount. That means, on average, 80% of the people with that score would have made the purchase anyway without being sent a $5 off promotion. You get some lift but it's not enough to cover the cost of the total discount dollars given.
Another way to look at it is like this: say you have a customer segment that spends on average $100 a month. You send them a $10 off coupon and that month they spend $100 again - minus the $10 off coupon. So you net $90 in spend and end up with less profits. You left $10 on the table.
On the flip side, when you see lower scoring segments making a ton of money, that is cold, hard as steel evidence that your retention program is working. Literally, this data represents money that was previously left on the table because controls did not spend it and the test group did.
Look at this, the top 10 most profitable scores in the numbers you sent:
Here are the 10 lowest incremental sales (highest subsidy cost) generators in the data:
So, why do some high scores generate profits and others losses? Why do some low scores generate profits and others losses?
We've discussed some of the potential reasons above. But this question really is unimportant. What is important is knowing you can generate $1.95 in sales per customer doing a certain promotion to 345's. And that you lose $1.16 in sales doing the same promotion to 532's.
Don't over-think the results, they simply are what they are. Over time, based on a deep understanding of your business, testing different promotions, and watching the consistent (by score) patterns of profitability provided by the RFM scoring, you will build a theory as to why people in certain scores behave as they do.
Here's a tip: if you must send discounts to high RFM scores with a tendency to create high subsidy costs, try "thresholding" the discount at the average ticket. In other words, if the average transaction of 532's is $50 a month, then send a "$5 off purchase over $50" discount. Better yet, customize the offer to the average transaction of each customer for the past 60 days or so.
By using the actual response and profitability behavior of your customers to target offers, you will learn over time what types of promotions drive the highest profits for each RFM score. Once you crack that code, you are on your way to driving serious profits!
That's it for this month's edition of the Drilling Down newsletter. If you like the newsletter, please forward it to a friend! Subscription instructions are top and bottom of this page.
Any comments on the newsletter (it's too long, too short, topic suggestions, etc.) please send them right along to me, along with any other questions on customer Valuation, Retention, Loyalty, and Defection here.
'Til next time, keep Drilling Down!
- Jim Novo
Copyright 2005, The Drilling Down Project by Jim Novo. All rights reserved. You are free to use material from this newsletter in whole or in part as long as you include complete credits, including live web site link and e-mail link. Please tell me where the material will appear.
What would you like to
the book with Free customer scoring software at: Learn Customer
Marketing Models and Metrics (site article
Get the book with Free customer scoring software at:
Learn Customer Marketing Models and Metrics (site article list)