A weekday model trained on 2010 records reduced log loss by 1.7% for 853 UK customer IDs at one retailer in January–November 2011

We tested whether past weekday counts add information about later shopping records from one UK gift retailer. We selected the prior count within 2010 and trained the final model on all 2010 records. The test used 6,384 shopping dates from 853 UK customer IDs in January–November 2011. Mean log loss fell from 1.78155 with a common weekday model to 1.75153 with personal counts, a reduction of 1.7%. The mean log score gain was 0.03002 nats per date. A customer bootstrap gave a 95% range of 0.01923–0.04159. Profiles exchanged between IDs with the same training count did not reach the observed gain in 1,999 draws. An exact score split added after the main test assigned 98.2% of this fixed model’s gain to Sunday versus other days. A separate model selected without Sunday records gave a smaller positive gain. The results concern recorded invoice dates from returning accounts at one retailer. They do not establish a purchase cause or predict whether an order will occur. The package includes input data, source hashes, code, all control draws, figures, and repeat instructions.

shopping dates · weekday profiles · chronological test · retail records · log score · customer identity