In addition to the many challenges of RFM already discussed, the segmentation puts up artificial barriers between segments. Some of these include:
- Let’s say someone is one of the people we talked about yesterday that gives every November or December. If s/he gave in November 2014, then again in December 2015, are you really going to consider them having “lapsed” in the middle?
- The distinction between frequency groups is artificial. As we discussed on Tuesday, Sandy, who gave you 100 gifts, and Miriam, who gave you two, are both considered multidonors for most RFM segmentations.
- The distinction between monetary value segments is artificial. Which donor would you prefer – a donor who donates $10 ten times per year or a donor who donates $50 once a year? RFM prefers the latter; I’m guessing you would prefer the former.
But how to do you create equivalencies among every different segment? Would you rather have a donor who gave $100 to an acquisition package six months ago or a loyal semi-frequent $20 donor?
The ideal would be to run a model with lifetime value as the dependent variable and your traditional RFM variables, plus as many of the ones that we’ve talked about this week, and determine what your actual drivers of value are.
But lifetime value, as you can tell from the name, takes a long time.
So let’s steal a rule-of-thumb model from the for-profit world. Connie Bauer first (at least first to my knowledge) proposed this in an influential 1988 Journal of Direct Marketing article called “A Direct Mail Customer Purchase Model.” Here, I’ve replaced purchases with donations; I think it works in our world with this replacement. To get the RFM score, you multiply these three things together:
- The reciprocal of recency of the last donation in months.
- Number of donations
- The square root of the total amount of donations the person has made.
There are a few things I like about this shorthand:
- There’s a reasonable equivalence between recency and frequency. Would you rather have someone who has given four gifts who gave their last gift a year ago or someone who has given two gifts and their last one was six months ago? These would be roughly equivalent in this model and that looks about right.
- It mitigates the artificial distinction between months. That 12-month versus 13-month difference that in a normal RFM analysis could be the difference between sending and not sending a communication? In this model, it’s about an 8% difference in scoring. Important, but not fatal.
- Because I’ve not seen the effect of the sheer numbers of gifts have a huge impact (once you get above a certain point) on retention rate, it seems intuitive that monetary value is a smaller factor than the other two.
There are some weaknesses. Donation amounts aren’t linear: if someone has given a $25 gift in the past, the odds that they will go from there to $26 to $27 is not likely. Some time periods, like a year, are somewhat magical, especially for one-gift-per-year, seasonality-focused donors. And in an ideal world, you would want more recent gifts weighed a bit more than more distant gifts. A donor’s behavior tomorrow will be more like their behavior last month than their behavior in 1988.
But given that, it’s an interesting look at the topic. I hope the week gives you the courage and the tools to take another look at your segmentation strategy and calculations. You’ll go nuts if you try all of these simultaneously, but conscious and continuous improvement can make huge differences in the long term.