RFM analysis
- Recency — days since the last purchase. Often the strongest signal: someone who bought recently is more likely to buy again.
- Frequency — number of purchases in a period.
- Monetary — amount spent in the period.
The classic approach is to give each customer a 1-to-5 score on each dimension (for example by quintiles) and combine the three scores: a 5-5-5 is a recent, frequent, high-value customer; a 1-4-5 is a customer who used to be worth a lot and has stopped, the first to win back.
From RFM to clustering
RFM is simple and readable, but the thresholds are set a priori. Clustering (for example with the k-means algorithm) uses the same variables — and others, such as average order value and channels used — to find the groups the data shows. The hard question is how many groups to make: too many is noise, too few hides differences. And sometimes the honest answer is that the customer base is homogeneous.
How to do it
1. Prepare the data
One customer per row, with purchase dates and amounts; better still, the orders file, one row per purchase. Merge customers who appear in more than one channel, or you’ll count them twice.
2. Calculate the variables
Recency, frequency, value, average order value, channels. Decide the reference period: twelve or twenty-four months, depending on how often people buy in your category.
3. Find the groups, and validate them
Whether they’re RFM scores or clusters, check that the groups are genuinely different from each other and stable: if they change a lot when you rerun the calculation, they aren’t real groups.
4. Measure who you’re losing
For each customer, how far behind their usual purchase rhythm they are: a customer who buys every month and has been silent for three is more at risk than one who buys once a year and has been silent for six. Add up the value of those slipping away: that’s the value at risk.
5. A plan for each group
Retain, grow, win back, let go. Each group has its own objective, channels and metrics.
The most common mistakes
- Demographic segments passed off as behavioral ones. Age doesn’t tell you who’s stopping buying.
- The same period for every category. Ninety days of silence is a lot for coffee, not much for a sofa.
- Too many groups. If you can’t give each one a different plan, there are too many.
- The average customer. An average doesn’t exist: better the real customer who best represents the group.
- Personal data where it isn’t needed. Behaviors and values are enough to segment: emails and names need protecting.
