customer-segmentation

Run the rfm subcommand with an analysis period and bucket count, for example segment rfm --period 90d --buckets 5. It scores customers on recency, frequency, and monetary value using configurable weights and assigns them to segments like Champions, Loyal, At Risk, Lost, and New.

Use the churn subcommand with a model type and confidence threshold, such as segment churn --model logistic --confidence 0.7. It scores each customer's churn risk from recency, frequency, monetary, support ticket, and browsing features, returning customers above the threshold.

You need at least 90 days of customer transaction history and access to order or sales data, typically via a PostgreSQL database set in SEGMENT_DB_URL. Lookalike audiences additionally require a seed segment of at least 5000 customers for acceptable quality.

Churn models degrade over time due to model drift, so retrain every 90 days and monitor AUC scores. Also review RFM bucket boundaries quarterly, since fixed thresholds may not fit all product categories.

Do not use it for loyalty points or tier management, individual customer service cases, or product-level analysis; those belong to loyalty-program, CRM, or assortment skills. Segments also do not prescribe actions, so always pair them with specific campaign strategies and A/B tests.