Customer Segmentation: Which Groups Behave Differently?

Question: Can two years of transaction history be converted into a trustworthy customer-level analysis table? Analysis: I audited 1,067,371 transaction lines, separated successful purchases from cancellations, and used a fixed 365-day observation window ending on 1 September 2011. Customers needed at least one successful purchase during that window. Fifteen features capture value, frequency, lifecycle, purchase cadence, product breadth, recent trend, and cancellation behavior. Finding: The final clustering population contains 4,335 customers. Missing purchase intervals are treated as expected single-order behavior, while extreme values are clipped rather than deleting customers. The following 90 days remain completely outside the clustering features.

Question: Which cluster count balances statistical quality, stability, and operational usability? Analysis: I evaluated K-Means solutions from two to eight clusters using silhouette, Davies–Bouldin, Calinski–Harabasz, perturbation stability, and minimum segment share. The operational choice was restricted to four to six clusters with no segment smaller than 3%. Finding: The five-cluster solution provides the best combined rank within the operational range. Its perturbation stability ARI is 0.974 and its smallest segment contains 9.5% of customers. The silhouette score is only 0.225, so the segments are presented as a stable operational simplification—not five strongly separated natural customer types.

Question: How do K-Means, Ward hierarchical clustering, Gaussian Mixture, and HDBSCAN compare on the same features? Analysis: Every algorithm uses the same clipped, transformed, imputed, and standardized feature matrix. I compare separation, cluster balance, and noise share, then use adjusted Rand index to measure agreement between K-Means and Ward labels. Finding: K-Means retains five operationally sized segments. Ward creates a 1.6% micro-segment, while HDBSCAN labels 68.3% of customers as noise. K-Means and Ward have moderate agreement (ARI 0.532). K-Means is selected because it is stable, reproducible, and can assign future customers to the nearest center; hierarchical clustering remains a structural cross-check.

Question: Do the cluster profiles translate into distinct, auditable customer strategies? Analysis: I profile each cluster in original business units and assign names through reproducible relative rules. Names summarize value, recency, tenure, recent activity, and cancellation behavior; they are operational labels rather than model-discovered personalities. Finding: The five groups are high-value loyal, stable repeat, new / high-potential, high-cancellation / low-engagement, and dormant risk. High-value loyal customers have median 365-day spend of GBP 3,828 and nine orders. Dormant-risk customers represent the largest group at 36.7%, while the high-cancellation group has a 50% median cancellation-invoice rate and should be investigated before promotional outreach.

Question: Do customers in different historical clusters behave differently during the following 90 days? Analysis: Future purchases were excluded from clustering and used only for external validation. I compare 90-day repurchase rates, order counts, spend, and Wilson confidence intervals across the five segments. Finding: High-value loyal customers repurchase at 87.1%, compared with 60.4% for stable repeat, 48.6% for new / high-potential, 35.3% for high-cancellation / low-engagement, and 27.9% for dormant-risk customers. This supports prioritization, but it does not prove that coupons or outreach will cause retention; that requires a randomized experiment with margin and treatment-cost data.