What is RFM segmentation in a charity context?

RFM scores each supporter on three dimensions: how recently they gave, how often they have given, and how much they have given in total or on average. The result is a small set of segments that lets a fundraising team treat distinct donor behaviours differently rather than sending the same ask to everyone.

Do I need a data scientist to run RFM?

No. A 3x3 scheme can be calculated in Excel from a transactions export, or built directly as saved lists or segments in Donorfy, Beacon, eTapestry, ThankQ and Salesforce NPSP. A data scientist is useful later if you move from RFM to predictive lifetime value or propensity models.

How often should I refresh the model?

For most small and medium charities, once a year is enough, ideally before the autumn appeal planning cycle. Larger charities with monthly direct debit programmes can refresh quarterly. The thresholds themselves should be reviewed annually, not the scores alone, because donor behaviour drifts over time.

When should we move beyond RFM?

When you can show that RFM cells consistently predict response and value, and you have at least two to three years of clean transaction history, it is worth piloting predictive models for lapse risk, next gift value or legacy propensity. Until then, RFM gives most of the operational benefit with a fraction of the complexity.

RFM Segmentation For Charity Databases Without Overengineering

Why 5x5x5 RFM grids fail on UK charity data, a 3x3 scheme you can run in Excel, Donorfy, Beacon, NPSP, eTapestry or ThankQ, the action per cell, an annual refresh routine, and how to test RFM-driven asks against broadcast appeals.

RFM is one of the oldest and most useful tools in direct marketing, and one of the most frequently mangled when fundraisers try to apply it to a charity database. The idea is simple: score every supporter by how recently they gave, how often they have given, and how much they have given, then treat the resulting segments differently. The trouble starts when teams reach for the textbook 5x5x5 grid and discover that charity data does not behave like a supermarket loyalty programme.

Where RFM came from, and why charities are not shops

RFM was formalised in the 1990s by direct mail catalogue businesses, who had millions of customers buying low-value goods many times a year. In that world, a 5x5x5 grid produces 125 cells, each with a healthy population, and the differences between cell behaviours are statistically obvious within a single quarter.

Charity files look nothing like that. A typical cash donor file has a long tail of one-off gifts, a smaller core of regular givers, and a handful of major donors whose individual gifts can distort the monetary axis on their own. Apply a naive 5x5x5 to that population and you get sparse cells, unstable scores, and a frequency dimension that is almost useless because most cash donors have given once or twice in their lifetime.

Why the textbook 5x5x5 fails on donor data

Three problems show up almost immediately when you try to run a 5x5x5 against a UK charity database.

Frequency is heavily compressed. For cash donors, the distribution is dominated by ones and twos, so quintile cuts produce groups that are not meaningfully different.
Monetary is skewed by one-off large gifts. A single capital appeal donation or in-memoriam gift can push a supporter into the top monetary band even though their typical behaviour is a £25 annual cheque.
Recency interacts with campaign cadence. If your last appeal was three months ago, almost everyone looks recent; if it was eighteen months ago, almost no one does.

The fix is not to abandon RFM. It is to design a scoring scheme that fits the shape of your data, not the shape of a 1995 mail order catalogue.

A pragmatic 3x3 scheme small and medium charities can actually run

A 3x3 grid on each axis gives 27 cells, which is usually enough resolution for a file of fifteen to a hundred and fifty thousand records. For files under fifteen thousand, drop frequency to two bands or collapse to a 3x3 RM grid and keep frequency as a flag.

Define the bands

Recency in months since last gift: 0 to 12, 13 to 24, 25 plus.
Frequency in lifetime gifts: 1, 2 to 4, 5 or more. For regular givers, treat any active direct debit as automatic top frequency.
Monetary in average gift value, not total: under £25, £25 to £100, £100 plus. Average is more stable than total and is less distorted by a single legacy or capital gift.

Build it in your CRM

In Donorfy, Beacon and eTapestry, the bands can be expressed as saved lists or selections from transaction queries. In Salesforce NPSP, a small set of formula fields on the Contact, plus a nightly batch to refresh the scores, is enough. ThankQ users can compute the scores in a SQL view and write them back to a custom attribute. If your CRM is fighting you, do it in Excel from a transactions export and upload the result as a tag, then automate later.

Use average gift value, not total. Total monetary is dominated by tenure, which double-counts recency and frequency and quietly hides genuine high-value behaviour.

What to do with each cell

A scoring model that no one acts on is just decoration. Every cell needs a default treatment, even if that treatment is "leave alone for now".

High recency, high frequency, any monetary: stewardship first, ask second. Thank, report back, and use these donors to test new propositions before broadcast.
High recency, low frequency, mid to high monetary: second gift conversion. The single biggest lift in lifetime value comes from turning a one-off donor into a two-gift donor within twelve months.
Low recency, high frequency, high monetary: lapsed major donor recovery. Personal contact from a fundraiser or trustee, not an email blast.
Low recency, low frequency, low monetary: suppress from print, keep on low-cost digital, and reappraise at the next refresh.
Mid cells across the board: this is where most of your file lives. Test message and channel variants here, because the volume is large enough to read results.

Most charities do not have a segmentation problem. They have a decision problem: too many segments, not enough agreed actions per segment.

The annual refresh routine

Set a fixed point in the year, usually six to eight weeks before the autumn appeal, and run the same three steps every time. Re-export the transaction history, recompute the band thresholds against the current distribution, and rescore the file. Compare the new cell sizes to last year and write a one-page note explaining any material shifts. If you skip the threshold review and only rescore, the model slowly stops matching your donor base.

When to graduate from RFM to predictive modelling

Predictive models are worth the investment when three things are true: you have at least two to three years of clean transactions, RFM cells are demonstrably predicting response and gift value, and you have a specific decision that a probability score would change, such as which lapsed donors to call or which warm prospects to invite into a legacy journey. Until those are true, a tidy 3x3 will out-perform a half-finished propensity model every time.

Test the impact, do not just believe in it

The honest way to find out whether RFM is earning its keep is to run a holdout. On the next appeal, take a random ten to twenty percent sample of mailable records and treat them as a broadcast control: same creative, no segmentation overrides. Send the remaining records the RFM-tailored version. Compare response rate, average gift and net income per record across the two groups. Repeat across two or three appeals before drawing conclusions, and keep the holdout the same shape each time so the comparison is fair.

If the RFM-driven version does not beat broadcast on net income per record after a couple of cycles, the model is wrong, the actions per cell are wrong, or both. That is useful information, not a failure. The point of segmentation is to spend less to raise more, and the only way to know whether it is doing that is to measure it.

Keep the model small enough that the fundraising team can describe it from memory, refresh it once a year, attach a default action to every cell, and test it against a broadcast control. That is most of the value of RFM, without any of the overengineering.