Counting cancellation records and same-day invoices raised repeat labels by day 30 for 553 of 3,565 customers at one retailer

We tested how event rules change repeat-purchase labels in UCI Online Retail II. The main cohort had 3,565 customers. Their first observed positive invoices were from March 2010 through September 2011. All customers had 90 full later calendar dates of follow-up. By the end of the 30th later date, 601 customers had a positive invoice on a later date. Counting later same-day positive invoices and cancellation records gave 1,154 positive labels. The rates were 16.86% and 32.37%. The difference was 15.51 percentage points. The descriptive 95% paired customer bootstrap range was 14.36 to 16.69 points. Of the 553 changed labels, 422 had cancellation evidence only, 92 had same-day purchase evidence only, and 39 had both. The difference was 14.61 points in the UK-only check and 15.03 points with a longer initial observation period. A separate XML parser checked every document. A separate SQLite implementation checked all 100,868 label rows across seven cases and four horizons. A full replay reproduced ten numerical and data files byte for byte. Broader rules must give at least as many labels. This study measures the size and source of that change in one record. Positive invoices do not prove completed purchases. The initial observation period does not prove that customers are new. The study does not identify customer intent or a causal effect. The package contains source data, provenance, code, labels, results, figures, and repeat instructions.

repeat purchasing · event definition · customer activity · cancellation records · measurement audit · Online Retail II · open data · reproducibility