RFM Analysis of Bank Clients

▶ Data-flow diagram — self-contained HTML, opens in a browser — no install, no build

Situation

The bank accumulates client transaction data, but marketing campaigns ran “flat”: the same offer for everyone. The task is to split the client base into homogeneous segments along three dimensions — Recency (when the last purchase happened), Frequency (operation count), and Monetary (total revenue) — and build a separate retention and growth strategy for each group.

Task

I needed campaigns to stop sending the same offer to every client, so I owned the split of the client base along recency, frequency, and monetary value, and the separate retention strategy that follows for each group.

Actions

Data: synthetic transaction history (seed = 42): 10,000 transactions → 9,492 after the anomaly drop, 1,983 active clients, period 2023-01-01 → 2024-03-30.

Analysis steps:

1. Cleaning & preparation — duplicates removed, rows flagged is_anomaly dropped.
2. RFM metric calculation per client:
   • Recency: days since last transaction
   • Frequency: number of operations over the period
   • Monetary: total revenue from the client
3. RFM scoring — each dimension scored in quartiles 1..4; segments are combo-based, not sum-based (so 444 and 233 no longer share a bucket).
4. Business-sense check — Lorenz curve and Gini coefficient, plus per-segment concentration_ratio.
5. Segment distribution visualization in Tableau with time and product filters.

Tools: Python (Pandas, Scikit-learn), SQL, Tableau.

Result

Six named segments emerged (quartile-score combinations):

• Champions (R≥4, F≥4, M≥4) — 126 clients, 6.4% of the base, driving 12.1% of revenue (concentration ratio 1.91).
• Loyal — loyal clients, just below Champions on all three axes.
• Potential Loyalists — recent with moderate frequency: upsell potential.
• At Risk — previously active but long inactive: need reactivation.
• Hibernating — 643 clients, 32% of the base, under-indexing on revenue (ratio 0.61).
• Need Attention — the remaining combinations.

Revenue concentration is measured, not asserted: Gini = 0.324, and the top 20% of clients hold 38.2% of revenue. At Risk + Hibernating ≈ 50% of the base — the main reactivation pool, described as concrete personas for targeting.

Limitations

The segmentation runs on a synthetic, reproducible extract (seed = 42) — the effect size here is not the point; the value is the scoring methodology and the concentration check. RFM is a rule, not a model: the score thresholds are set by the analyst and would need revision on a different dataset.

Documentation

• GitHub → rfm-analysis-of-bank-clients