In an RFM customer segmentation, what do recency, frequency and monetary value each measure?

## The three inputs RFM summarises each customer with three numbers computed over a fixed observation window (say the last twelve months): - **Recency** - time since that customer's most recent purchase, usually in days. **Lower is better.** This axis is inverted relative to the other two, which is the single most common confusion. - **Frequency** - how many purchases the customer made inside the window. Higher is better. - **Monetary** - how much they spent inside the window, either total or average per order. Higher is better. State the window explicitly. "Four purchases" means nothing until someone says four purchases *since when*, and two analysts using different windows will produce two incompatible segmentations of the same customers. ## Why these three Each answers a question the other two cannot. Two customers with identical spend can be a lapsed big spender and a steady regular; recency separates them. Two customers with identical recency can be a habitual weekly buyer and a one-time impulse purchase; frequency separates them. And frequency alone is blind to whether the habit is worth anything, which monetary supplies. Together they cover the alive/habitual/valuable dimensions of a transactional relationship using nothing but a sales table. ## Two ways it is used **The score variant.** Rank customers into quintiles on each axis, 1 to 5, and concatenate: a `555` is recent, frequent and high-spending; a `155` is a former big spender who has gone quiet. Because it uses ranks rather than raw amounts, it is immune to the heavy right tail of spend and every tier is defined by a rule anyone can reproduce next quarter. Its weakness is that quintile cuts are arbitrary and 125 code combinations are too many to act on, so codes get grouped into a handful of named tiers. **The clustering variant.** Feed the three numbers to a clustering method and let the partition fall out. More flexible, and it will find groupings the quintile grid cuts through, but the boundaries are no longer explainable as a rule and they move whenever the model is re-fitted. ## Naming the result The output only becomes useful when each group carries a name a marketer can repeat. A retail base typically resolves into tiers like **Champions** (bought recently, several times, top-decile spend - often a small share of customers carrying a large share of revenue), **At-Risk** (used to buy often and spend well, but nothing recently), and **Hibernating** (long gone, low frequency, low value). Those three names imply three different actions, which is the point of doing the exercise at all. ## What RFM does not carry - **Margin.** A high-monetary customer who only ever buys discounted clearance stock can be unprofitable. Revenue is not value. - **Product mix and affinity.** RFM cannot tell a customer who buys one category from one who buys across the catalogue. - **Tenure and acquisition channel.** A `555` who joined last month and a `555` of ten years look identical. - **Seasonality.** Someone who reliably buys once a year in December looks lapsed every July. A window shorter than the natural purchase cycle manufactures fake churn. - **Subscription businesses.** With a flat monthly price, monetary value is nearly constant across subscribers, so it separates almost nobody, and "purchases" are automatic renewals that measure billing rather than engagement. ## Choosing inputs when RFM does not fit For a streaming service the useful inputs are behavioural: sessions per week, genres watched, and cancel-page visits. Prefer behavioural inputs over demographics such as age band, country or signup channel for two reasons. First, demographics are weakly related to what people actually do - plenty of 25-year-olds behave like the average 55-year-old subscriber. Second, behaviour is the thing an intervention can move and the thing that changes first when a customer is about to leave; a birth year cannot be influenced and never updates. That does not make demographics useless. They earn their place as **profiling** variables - describing the segments after the partition exists, and often supplying the name and the channel - rather than as clustering inputs. ## Practical cautions Remember the inverted recency axis when you sort, chart or name anything. Expect monetary value to be heavily right-skewed, so a handful of very large accounts sit far from everyone else; the quintile-ranked variant sidesteps that by using ranks rather than raw amounts. And be clear about what RFM is: it *describes* the current state of a customer base. It is not itself a forecast, and a segment name like At-Risk is a hypothesis about the future, not a measurement of it.