RFM Analysis

RFM segmentation ranks customers on three facts from their purchase history. Recency is how long ago they last bought. Frequency is how many times they bought. Monetary value is how much they spent. Each fact gets a score, and customers with the same three scores form a segment. The method also goes by RFM analysis and the RFM model.

Where does the RFM model come from?

RFM is a direct marketing technique. IBM's SPSS manual calls it a way to find the existing customers most likely to respond to a new offer. The manual adds that it is common in direct marketing (IBM). Fader, Hardie and Lee write that the direct marketing literature commonly sums up past buying this way (Journal of Marketing Research, 2005).

The order of the letters is a ranking. IBM's manual names recency as the most important factor, then frequency, then total spend (IBM). Fader, Hardie and Lee studied customers who first bought from the online music store CDNOW in early 1997. Future value varied most with recency and almost as much with frequency. It varied least with monetary value. They tie this to the view that recency usually predicts best, which is why the name puts recency first (Fader, Hardie and Lee).

How do you calculate RFM scores, step by step?

1. Pick the data and the window. You need a customer ID, a date and an amount for each transaction (IBM). Use a window that covers a normal buying cycle, such as 12 months.
2. Work out three raw values. For each customer, count the days since the last purchase and the number of purchases in the window. Then add up their spend. Spend is typically the total, though the mean or the largest purchase also works (IBM).
3. Rank and bin each value. Sort customers on each value and cut the list into five equal groups. The top fifth scores 5 and the bottom fifth scores 1. For recency, the most recent buyers score 5 (IBM).
4. Choose independent or nested bins. Independent bins score each value on its own, so a frequency score of 5 means the same for every customer. Nested bins rank frequency inside each recency group. That spreads customers more evenly across cells, but the scores get harder to read (IBM).
5. Settle ties. Many customers share a value, such as one purchase or the same last date. Putting ties in the same bin keeps scoring consistent but can make bins uneven. In IBM's example, if 500 of 1,000 customers last bought on the same day, half get a recency score of 5 instead of a fifth (IBM).
6. Combine the scores. Write the three scores side by side. Five bins each give 125 cells, from 111 to 555 (IBM). Do not add the digits, because 511 and 115 both sum to 7 and would look the same.
7. Name the groups and give each one a job. Merge cells that share an action into a few segments, each named for what the program will do.
8. Rescore on a schedule. A score is out of date as soon as the customer buys again. Programs that act every day need dynamic segments that rescore as members act.

What does an RFM analysis look like with real numbers?

For illustration, a grocery program scores a year of member purchases. Its cutoffs come from its own data:

• Recency: 5 for 0 to 14 days since the last purchase, 4 for 15 to 30, 3 for 31 to 60, 2 for 61 to 120 and 1 for more than 120.
• Frequency: 5 for 24 or more purchases, 4 for 12 to 23, 3 for 6 to 11, 2 for 2 to 5 and 1 for a single purchase.
• Monetary value: 5 for $2,000 or more, 4 for $1,000 to $1,999, 3 for $500 to $999, 2 for $200 to $499 and 1 for under $200.

Members A and B spend about the same, yet they need opposite treatment. A gets recognition, since a discount would pay for visits A already makes every week or two. B has been silent for 95 days and needs a win-back campaign now.

How do loyalty programs use RFM segmentation?

A loyalty program makes RFM more accurate. A member ID ties every purchase to one person across every channel. Without it, a shopper who pays with two cards looks like two light buyers. With it, frequency and spend are counted in full. Programs then use the cells in four places.

• Offers. Cells decide which members get an offer and how rich it is. A discount to a member who buys every week pays for visits that were coming anyway.
• Win-back. High frequency with a low recency score flags valuable members who have stopped. They are the first call for a win-back offer.
• Tier design. Tier thresholds set near the natural breaks in spend and frequency give members a stretch they can reach.
• Program health. Rescore monthly and count the members who move up or down a cell. More moving up than down means the program is working.

Decide early how redemptions count. A clean rule: redemption visits count toward frequency, and monetary value stays cash spend, so rewards do not inflate a score.

What are the limits of RFM analysis?

• It looks back. Fader, Hardie and Lee note that scoring models built on these measures predict the next period. Lifetime value needs every period after that (Fader, Hardie and Lee).
• The link to future buying bends. The same study found a highly nonlinear link between past recency and frequency and future purchases. Among customers who had not bought in a long time, frequent past buyers were likely to have lower future potential than light buyers. The authors call this the increasing frequency paradox.
• Quintiles move. A 5 means the top fifth of this period, so the spend behind a 5 shifts as the base changes. Fix the cutoffs when you compare scores over time.
• Spend ignores margin. Monetary value counts revenue and leaves out margin and the cost to serve. A member who buys only on promotion can score high and still lose money.
• It counts as profiling. Under the EU's GDPR, profiling is the automated use of personal data to judge or predict things about a person (GDPR Article 4). People can object to its use for direct marketing at any time (GDPR Article 21).

The fix is to keep RFM as the base and add a forward-looking model. Fader, Hardie and Lee built one that forecasts lifetime value from each customer's RFM values alone. In their model, no other data about customer behavior is required (Fader, Hardie and Lee). The customer lifetime value guide covers that next step, and the customer segmentation guide sets RFM beside the other segment types.

How GRAVTY handles RFM segmentation

GRAVTY is the Loyalty Juggernaut platform. Member 360 pulls a member's behaviors and transactions into one view, household included. So the purchase history RFM needs sits in one place. Segments are written in GRAVTY Visual Rules, a patented no-code rules engine, and they read live member data and model scores.

Scores built by an outside data team slot in as member attributes. That makes them segments a campaign can target, with no rebuild. Rules run in a simulation before they go live, and the team ships them without an IT ticket. Agentic AI Compass compares results across any segment in plain language. GRAVTY runs 500M+ members in production.

Frequently asked questions

What is a good RFM score?

In a five-bin scheme, 555 is the highest score, and IBM's manual treats the highest combined scores as the customers most likely to respond to an offer (IBM). Below the top, a good score depends on the action. A 255 is weak on recency, yet it marks a big spender worth a fast win-back.

What data do you need for RFM analysis?

You need a customer ID, a date and an amount for each transaction (IBM). Data already summed per customer works as well, as long as it holds the last purchase date, the purchase count and a spend figure for each one.

How many RFM segments should a loyalty program use?

Use one segment for each distinct action, and merge cells that would get the same one. A program with eight offers needs at most eight segments, even though five bins on each score make 125 cells.

Is RFM still useful when you have machine learning?

Yes, because in Fader, Hardie and Lee's lifetime value model, a customer's RFM values are the only purchase data needed (Fader, Hardie and Lee). RFM also gives any model a simple baseline to beat, and a finance team can follow it.

What is the difference between RFM and customer lifetime value?

RFM describes how a customer bought in the past, while customer lifetime value estimates what they will be worth in the future. Fader, Hardie and Lee link the two with iso-value curves. Each curve groups customers with different histories but the same expected future value (Fader, Hardie and Lee).