RFM is the best return on effort in customer analytics. It needs one table of orders and an afternoon. I have strong opinions on how it gets misused, so I will mark them as opinions as we go.
Three columns, one row per customer
Take the order history and collapse it to one row per customer. Recency: days since the last order. Frequency: number of orders in a window, say the last 12 months. Monetary: total spend in the same window. That is the whole data requirement.
Six customers from a hypothetical home-décor store show the idea. A: last order 5 days ago, 8 orders, ₹14,400. C: 200 days ago, 6 orders, ₹9,000. D: 30 days ago, 1 order, ₹700. F: 3 days ago, 12 orders, ₹21,600. B: 12 days, 3 orders, ₹4,500. E: 400 days, 2 orders, ₹2,500. Total spend across the six is ₹52,700. A and F both average ₹1,800 per order, but F buys one and a half times as often. C has a respectable history but has gone quiet for months. D is new, or a one-off. Spend alone would rank C above B; recency says B is the one still talking to you.
Turning raw values into scores
Raw values are on different scales, so convert each to a score from 1 to 5. The usual way is quintiles: sort all customers by the metric, cut into five equal-sized groups, and give the best group a 5. For recency the best group is the smallest number of days, so the direction flips. Each customer ends up with a code like 5-4-5, or 1-1-2.
Opinion: quintiles are a default, not a law. If most of your customers bought exactly once, the frequency quintiles collapse, because the bottom three groups all have frequency 1 and the tool will split ties arbitrarily. Look at the distribution first. When 70% of customers have one order, use three bands for frequency (1, 2 to 3, 4 or more) and be honest that it is coarse. Business-defined cut-offs beat statistical ones when you have a reason, such as “lapsed means no order in 180 days”.
The other trap: scoring with all customers in one pool when you sell both a ₹300 consumable and a ₹30,000 appliance. Monetary then just recognises who bought the appliance. Score within product families or channels, or at least check that the monetary score is not simply a proxy for category.
Segments are the point, not the scores
125 combinations are useless to a marketing team. Group them into a handful of segments with a clear action each.
- Champions: high on all three. Early access, referral asks, no discounts needed.
- Loyal but cooling: high frequency and monetary, recency slipping. A reminder or a “we missed you” note, perhaps a small incentive.
- New: high recency, frequency of one. The second order is the most valuable event in the whole relationship, so nurture it.
- At risk or lapsed big spenders: low recency, high monetary. Worth a personal or higher-cost win-back effort.
- Lost, low value: low everywhere. Stop spending on them, or one cheap reactivation attempt and then suppress.
Customer C in the example lands in the “at risk” bucket. Customer E, with 400 days of silence and low spend, is probably lost. Different money should go to each, and that is the entire value of the exercise.
Where RFM stops being enough
RFM describes the past. It does not tell you who will respond to an offer; a champion may have bought anyway, so discounting them wastes margin. Uplift thinking handles that question, and RFM is a good first filter before it, not a substitute.
It also assumes a repeat-purchase pattern. For a business where customers naturally buy every three years, such as furniture or insurance, a recency of 400 days is normal, not alarming. Calibrate recency to the natural purchase cycle of the product. And rerun the scores regularly: a customer’s code is a snapshot, and a segment that is not refreshed monthly goes stale quickly.
My last opinion: store the segment history. If you only keep today’s segment you cannot ask the useful question, which is how many champions of last year are still champions, and what did the ones who slipped have in common.
Figures and names in the examples are illustrative.
Leave a Reply