Segmentation Approaches
Pre-defined vs data-driven segmentation
Pre-defined segmentation: Segments are defined by business rules before analysis.
- Examples: pricing plan, geography, acquisition channel, industry vertical
- Advantages: immediately actionable, stable over time, explainable to stakeholders
- Disadvantages: may not reflect natural groupings in the data
Data-driven segmentation (clustering): Segments are derived from patterns in the data.
- Examples: k-means on behavioural features, RFM (Recency, Frequency, Monetary) quartiles
- Advantages: uncovers non-obvious groups
- Disadvantages: requires interpretation; segments change as data changes
For most business analyses, start with pre-defined segments. Use data-driven approaches when the goal is to discover segments, not confirm existing ones.
RFM segmentation
Recency, Frequency, and Monetary value — a standard customer segmentation for transactional businesses.
| Dimension | Definition |
|---|---|
| Recency | Days since last purchase |
| Frequency | Number of purchases in a period |
| Monetary | Total or average spend |
Score each dimension on a 1–5 scale (5 = best). Champions are 5-5-5; At Risk are 1-1-1.
Standard RFM segments:
| Segment | RFM pattern |
|---|---|
| Champions | High R, F, M |
| Loyal customers | High F, medium M |
| Potential loyalists | Recent, low F |
| At risk | Low R, formerly high F/M |
| Hibernating | Low R, low F |
| Lost | Very low R, low everything |
Index scoring
Index = segment metric / overall metric × 100
Interpretation:
- 100 = in line with average
- 120+ = over-indexing (20% above average)
- 80 or below = under-indexing (20% below average)
Index scores normalise for segment size and allow comparison across metrics of different scales.
Segmentation pitfalls
Simpson's paradox: A trend in the aggregate can reverse within every segment. Always break down before concluding.
Over-segmentation: Too many small segments produce noise, not signal. Segments with < 30 observations are generally not reliable.
Confounding with size: A segment that is 80% of total volume will always explain most of the movement. Look at rate changes, not just absolute contributions.
Survivorship bias: Analysing only active or paying customers ignores the population that churned or never converted.
Segment profile template structure
For each segment, document:
- Label — a business-meaningful name (not "cluster 3")
- Size — N and % of total
- Key characteristics — 3–5 metrics where this segment differs from the average
- Index scores — for primary metrics
- Hypothesised need or behaviour — why this segment behaves differently
- Recommended action — what product, marketing, or CS intervention makes sense