RFM Segmentation: Method, Assumptions & Examples

1. Introduction

1.1 Why Behavior-Based Segments

Demographic attributes say little about whether a customer will buy again. Order history shows what customers actually did: how recently, how often and how much. RFM turns that history into three numbers per customer that can be ranked and combined.

1.2 What This Paper Covers

• The definitions of recency, frequency and monetary value
• Scoring approaches and how to handle skew and ties
• A reproducible holdout test of whether the scores predict later purchases
• Guidance on segments, windows and limits

2. Background

2.1 The Three Measures

• Recency (R): days since the customer's most recent order.
• Frequency (F): number of orders in the analysis window.
• Monetary value (M): total spend in the window (average order value is an alternative).

RFM comes from direct marketing practice (Hughes, 1994), and later work (Fader, Hardie and Lee, 2005) shows how recency and frequency relate to probabilistic models of customer purchasing.

2.2 Scoring Approaches

Quintile scoring splits customers into five equal groups on each dimension, scoring 1 to 5 (recency scored so that the most recent customers get 5). A customer in the top group on all three scores 555; there are 125 combinations, usually grouped into a handful of named segments. Quartile scoring uses four groups and 64 combinations. Weighted scoring combines the three scores with weights chosen on business grounds or from a predictive model. Threshold scoring uses absolute cut-offs, such as ordered in the last 30 days, five or more orders and over a spend amount, so segments do not shift as the customer base changes.

3. Approach

3.1 Window and Data

Choose an analysis window that matches the purchase cycle: long enough to see infrequent buyers, short enough that old behavior is not mixed with current. Transactions need reliable timestamps, a customer identifier that is consistent over time, and values on one currency basis. Returns, test orders and duplicates should be handled before scoring.

3.2 Skew and Ties

Order counts and spend are right-skewed, and most customers in many businesses have only one or two orders. Cutting into quintiles by value then runs into many identical values, so the cut can fail or the bins come out unequal. Ranking the customers first (breaking ties by order) makes equal-sized bins, though customers with identical behavior can then land in different bins; document the choice. For spend, a log transform or extra bins at the top separate big spenders who would otherwise share a single top bin with moderate ones.

3.3 Segments

Name segments by the behavior they describe, for example: top on all three (best customers), high on frequency and low on recency (lapsed regulars), recent with one order (new customers). Keep the names and rules written down, because they will be reused in campaigns and reports.

3.4 Validate by Holdout

Compute the scores using data up to a cut-off date, then compare the share of each score group that bought in a period after the cut-off. If the scores do not separate buyers from non-buyers on your data, they do not help for targeting, whatever the segment names suggest.

4. A Reproducible Example

The code below simulates 4,000 customers with different steady purchase rates and basket sizes, scores the 2,752 who ordered in the 540-day history window, and checks who bought in the following 90 days.

import numpy as np
import pandas as pd

rng = np.random.default_rng(8)
N, DAYS_HISTORY, DAYS_AHEAD = 4000, 540, 90

# Synthetic customers: each has a purchase rate (orders per year) and a typical basket value
rate = rng.gamma(shape=0.8, scale=2.0, size=N)               # skewed: most buy rarely, a few buy often
basket = rng.lognormal(mean=3.8, sigma=0.6, size=N)
rows = []
for cust in range(N):
    n_orders = rng.poisson(rate[cust] * (DAYS_HISTORY + DAYS_AHEAD) / 365)
    day = rng.integers(0, DAYS_HISTORY + DAYS_AHEAD, size=n_orders)
    value = basket[cust] * rng.lognormal(0, 0.3, size=n_orders)
    rows += list(zip([cust] * n_orders, day, value))
orders = pd.DataFrame(rows, columns=["customer", "day", "value"])

history = orders[orders.day < DAYS_HISTORY]
future = orders[orders.day >= DAYS_HISTORY]
rfm = history.groupby("customer").agg(last=("day", "max"), frequency=("day", "size"), monetary=("value", "sum"))
rfm["recency"] = DAYS_HISTORY - rfm["last"]                      # days since last order
rfm["bought_next_90d"] = rfm.index.isin(future.customer.unique())

def quintile(series, ascending=True):
    """Score 1-5; rank first so tied values (many one-order customers) do not collapse the bins."""
    return pd.qcut(series.rank(method="first", ascending=ascending), 5, labels=[1, 2, 3, 4, 5]).astype(int)

rfm["R"] = quintile(rfm["recency"], ascending=False)             # most recent = 5
rfm["F"] = quintile(rfm["frequency"])
rfm["M"] = quintile(rfm["monetary"])

print(f"customers with at least one order in the history window: {len(rfm):,}; bought in the next 90 days: {rfm.bought_next_90d.mean():.1%}")
print("share who bought in the next 90 days, by score (1 = lowest, 5 = highest):")
print(pd.DataFrame({dim: rfm.groupby(dim).bought_next_90d.mean() for dim in "RFM"}).round(3).to_string())
champions = rfm[(rfm.R >= 4) & (rfm.F >= 4) & (rfm.M >= 4)]
at_risk = rfm[(rfm.R <= 2) & (rfm.F >= 4)]
print(f"R,F,M all >= 4: {len(champions)} customers, {champions.bought_next_90d.mean():.1%} bought next; "
      f"R <= 2 with F >= 4 (lapsed frequent buyers): {len(at_risk)} customers, {at_risk.bought_next_90d.mean():.1%} bought next")

Output:

customers with at least one order in the history window: 2,752; bought in the next 90 days: 35.6%
share who bought in the next 90 days, by score (1 = lowest, 5 = highest):
       R      F      M
1  0.218  0.194  0.214
2  0.278  0.233  0.280
3  0.325  0.296  0.284
4  0.495  0.418  0.425
5  0.466  0.641  0.579
R,F,M all >= 4: 579 customers, 60.1% bought next; R <= 2 with F >= 4 (lapsed frequent buyers): 158 customers, 41.1% bought next

In this simulation frequency separated future buyers best, spend came next and recency least, which is expected here because every simulated customer buys at a constant rate: a long gap tells you little beyond the rate. Real businesses with churn or seasonal habits often show recency mattering more. Which dimension leads is a property of the data, so the same holdout check on your own orders is worth more than any rule of thumb. The combined groups were sharper: customers scoring 4 or 5 on all three dimensions bought at 60%, and lapsed frequent buyers (recency 2 or lower, frequency 4 or higher) at 41%, against 36% overall.

5. Practical Guidance

6. Conclusion

RFM is a simple, explainable way to rank customers from order history. Its usefulness is empirical: in the simulation here the scores separated future buyers clearly, with frequency leading, and on a real customer base the result can differ. Score with rank-based bins, test against a later period, and build campaigns on the dimensions that predict in your own data.