RFM (recency, frequency, monetary)

What the three dimensions actually measure

Recency is the number of days since the customer's last order. It is the strongest of the three for most retail catalogues, for a mundane reason: buying behaviour is habitual and habits decay, so the best available evidence that someone will buy from you next month is that they bought from you last month. A customer who ordered three weeks ago is in a different state from one who ordered fourteen months ago, regardless of how much either has spent historically.

Frequency is the number of orders in a defined lookback window, and the window is a decision rather than a detail. "Orders ever" rewards tenure and permanently overstates customers who were active years ago; a rolling 12 or 24 months describes the customer as they are now. Pick one, state it, and do not change it quietly, because frequency scores are not comparable across different windows.

Monetary is total spend in the same window — sometimes average order value instead, which answers a different question. Total spend correlates with frequency, so the two scores partially duplicate each other; average order value is more independent and more useful when you want to separate "buys often, buys cheap" from "buys rarely, buys large". Both have the same blind spot, which is that they measure revenue rather than margin.

Scoring by quintile, worked through

The standard method is to rank the customer base on each dimension and cut it into five equal groups. Take an example book of 10,000 customers with at least one order in the last 24 months. Sort by days since last order, take the 2,000 most recent, and score them R = 5; the next 2,000 get R = 4, and so on down to R = 1 for the 2,000 who have been quiet longest. Repeat independently for frequency and monetary value. Each customer ends up with three digits, from 555 down to 111.

Quintiles are used rather than fixed thresholds because they are self-calibrating. A catalogue where the median gap between orders is three weeks and one where it is nine months both produce a sensible R = 5 group, and neither needs a hard-coded rule that goes stale as the business changes.

The table below shows what those cut points might look like for the example book. The important feature is not the numbers, it is what happens in the frequency column.

The tie problem, and how to handle it

Frequency of 1 appears twice in that table, in the bottom two quintiles, and in a real ecommerce book it usually spans three or four of them. Most customers of most brands have ordered exactly once. There is no way to cut a column that is 70% identical values into five equal groups, and every tool that tries produces arbitrary splits where two customers with identical behaviour land in different segments.

The workable approaches are to rank with ties and accept unequal bucket sizes, which keeps the scores honest at the cost of a lopsided grid; to score frequency on explicit thresholds instead of quintiles — 1, 2, 3, 4–6, 7 or more — which is readable and stable; or to drop frequency to a binary "has repeated" flag for catalogues where repeat purchase is genuinely rare. The one thing not to do is let a percentile function silently split the ties and then present the result as though the segments differ.

The same caution applies to the monetary dimension for a different reason: it counts revenue, not contribution. A customer whose entire history was bought on 40% off scores identically to one who paid full price for the same total, and only one of them is worth acquiring more of. If you can compute contribution margin per customer, use it in place of spend — it changes which customers appear in the top segments more often than people expect.

The segments the scores produce, and the action each implies

The scores are only useful once they are collapsed into a handful of named groups that carry an action. Six covers most catalogues. The patterns below are the common definitions; adjust the boundaries to your own distribution rather than importing someone else's grid.

Where RFM stops working

Subscriptions break it structurally. Every active subscriber has a recent order by definition, because the billing system created one, so recency measures the billing cadence rather than the customer's intent and every active subscriber scores R = 5 whether they are delighted or have forgotten they are enrolled. Frequency degenerates into tenure for the same reason. For a subscription book, replace recency with days since the last voluntary action — a login, a skip, a swap, an email open, an added item — and replace frequency with something that carries information, such as the number of skips or the payment-failure history.

Low-frequency, high-ticket catalogues break it differently. For mattresses, furniture or appliances, nearly every customer has a frequency of 1, and a low recency score is the expected state rather than a warning: somebody who bought a sofa eighteen months ago is behaving exactly as a sofa customer should. Scoring them as "at risk" and mailing them win-back offers is a category error. Here the useful axes are category, order value and referral behaviour, not RFM.

The general limitation is that RFM is descriptive and backward-looking. It tells you what a customer has done, which is a good proxy for what they will do next and no explanation of why. It cannot see that a cohort is lapsing because a formulation changed or because delivery times slipped, so it will faithfully route those customers into a win-back segment and offer them a discount to solve a problem that a discount does not solve.

None of that is an argument against it. RFM costs one SQL query, needs no model and no vendor, and it is a large improvement over sending the same message to everyone. Just do not let it be the only lens, and re-run it on a schedule — scores computed once and frozen become steadily more wrong, because recency changes for every customer every day whether they act or not.