This is a fully synthetic worked example for a customer-segmentation article. Use only customers.csv, order_lines.csv and refunds-v1.csv; keep all source IDs. The data contain no real people or contact information. Name the project "Customer Segmentation — Synthetic Worked Example".

Use snapshot 2026-10-01 00:00 UTC and purchase window 2025-10-01 inclusive to 2026-10-01 exclusive. Do not use today's date. Qualifying purchases are completed_paid orders with a valid roster customer, consistent order header fields, positive merchandise cents, and purchase before snapshot. For exact repeated line IDs/fields retain the lowest numeric source_row. A conflicting line ID excludes every affected order; inconsistent order customer/date/status, unknown/blank customer IDs and nonpositive orders are excluded and logged. Preserve legitimate separate merchandise lines within an order.

For refunds: deduplicate truly identical refund IDs; quarantine conflicts. Require a positive refund linked to a qualifying observed order, posted on/after purchase and before snapshot. If candidate refunds total more than the order's value, quarantine all candidates for that order. Keep rejected refunds and customer values needing repair visible. M follows orders inside the purchase window, not a separate refund-posting window; accepted refunds for older orders are reported separately.

Calculate one row per roster customer. R is UTC days since latest qualifying observed purchase before snapshot, blank if no qualifying history. F counts DISTINCT qualifying order IDs inside the purchase window. M is in-window merchandise cents less accepted refunds linked to those in-window orders, shown in USD. A completed fully refunded purchase still contributes to R/F. Separate no-history records from eligible buyers.

Apply FIRST matching buyer rule: F0 -> Historical outside window; otherwise M0 -> Zero-net refund review; otherwise R<=90/F>=2/M>=50000 cents -> Recent repeat higher-value; R>90/F>=2/M>=50000 -> Lapsed repeat higher-value; R<=90/F1 -> Recent one-order; R<=90 -> Other recent; remaining -> Other lapsed. These are illustrative business rules, not percentile scores, churn predictions or campaign results.

Build an inspectable customer review with R/F/M, qualifying-history status, group, source evidence and a sensible next check; save a downloadable CSV. Show segment counts/value totals and source exclusions, and reconcile customers, distinct orders and money from orders through customers to groups. Keep no-history records separately; do not call older buyers with F0 "no history". Include a clear segment summary view and source-checkable customer rows. Do not send campaigns or alter external accounts.