UK invoice records from one retailer show same-day clusters and fewer 1–7-day gaps between shopping days under calendar controls in 2010–2011
We test the timing of positive invoice records from one UK online retailer. We use UK customer records from January to November in 2010 and 2011 in UCI Online Retail II. There are 15,438 and 14,623 invoice events. In each year, we first count return gaps between invoices. We then merge all invoices from the same customer and date into one shopping day. Calendar controls keep each customer’s monthly count and each store day’s total fixed. A second control also keeps each customer’s weekday counts fixed within each month. Under this second invoice control, 1,701 and 1,303 same-day gaps are 2.12 and 1.95 times their exact expected counts. After daily merging, 47.7% and 46.3% of gaps within a month are 1–7 days. The corresponding fixed-margin day controls give means of 58.3% and 58.6%. The differences are 10.5 and 12.4 percentage points. Removal of the most frequent 1% of buyers leaves this day-level deficit. Thus same-day invoice clusters and wider gaps between shopping days occur in the same records under two specified conditional tests. The tests use different event units and constraints. They do not identify a causal effect of daily merging. Invoice times can reflect store operations, and many customers are wholesalers. The two periods share customers and are not independent samples. The package contains the original data, all random draws, code, checks, figures, and offline repeat instructions.
human dynamics · purchase timing · event definition · retail data · calendar controls · conditional randomization · Curveball
References
7- usesdoi:10.24432/C5CG6D
We use the original Online Retail II workbook and retain its CC BY 4.0 attribution to Daqing Chen.
- doi:10.1287/mksc.2014.0873
This study shows why purchase clustering can matter for customer value models. We do not test its predictive claim.
- doi:10.1140/epjds/s13688-018-0133-0
This work describes individual and collective temporal shopping profiles. Our contribution is a conditional audit of invoice and shopping-day records.
- doi:10.1088/1367-2630/14/1/013055
This work separates calendar rhythms from bursts in human activity. It motivates explicit calendar controls.
- usesdoi:10.1038/ncomms5114
We implement Curveball trades for binary customer-by-date matrices with fixed row and column totals.
- usesdoi:10.1103/PhysRevE.91.042812
This work gives the uniform stationary distribution of Curveball. Our chain retains failed and self trades as no-change steps.
- doi:10.1103/PhysRevE.94.039902
Erratum to the Curveball uniform-sampling proof.
Cited by
5 publicationsPublic catalog entries on Sciscape.
- Weekly cycles and weekly activity trends do not remove inter-event burstiness for 24,231 actors in 21 activity datasets under cross-fitted Poisson rate models
Related SCI test of event clustering under calendar controls in retail invoices.
- usesProduct overlap on consecutive shopping days differed by -0.48 and +0.09 Jaccard points from quarter order controls at one UK retailer in 2010 and 2011
Its invoice audit and timing results motivated daily merging. Our independent source extraction gives the same annual invoice and customer-day counts before the product-code rule.
- usesAdding distinct customer counts changed product pair forecast loss from 0.276025 to 0.276024 at one UK retailer in 2011
The study informs the invoice event rules. Its invoice totals provide a separate extraction check. The new analysis tests product pairs and distinct customers.
- supportsA weekday model trained on 2010 records reduced log loss by 1.7% for 853 UK customer IDs at one retailer in January–November 2011
Prior SCI analysis shows why invoice and shopping-date definitions and calendar controls matter in this source. Our target is future weekday score, rather than within-month gap counts.
- usesCounting cancellation records and same-day invoices raised repeat labels by day 30 for 553 of 3,565 customers at one retailer
Its finding of same-day invoice groups motivates this label audit. We reuse the retained raw source. We write new extraction and analysis code.